Rumours and claims to verify

Claims made on the cmpunlocker Discord that nobody has reproduced, and claims that were checked and turned out to be false. Each entry gives the claim, who made it and when (in the credit line), what would settle it, and the board card if there is one. Where another page already covers the claim with its sources, the entry is short and links there. Older history (July to early October 2026) sits in each section next to the October entries.

• Theory on this page: a claim with no evidence posted and no second person.
• Contested: people disagree. Both sides are listed in the order they were posted, with no verdict.
• Reported: one account of a claim, or one check that showed a claim to be false.
• Warning: a fake or a sales pitch that has already cost buyers money, or a claim that can cost you a card if you act on it.
• To settle one, post the evidence on Discord with the card variant, driver and method, and move the card on the board. A settled claim moves to the page it belongs on.

Firmware and VBIOS claims appear here only as claims. This wiki does not carry keys, decryption or ways around NVIDIA's firmware protection.

Memory

8 GB card at 80 or 96 GB

On 2026-10-09 @ob88594 posted a Xianyu listing photo of cards marked 80G and 96G, whose CUDA count matched 128 SMs. His informant said they were 8 GB cards; later he added that the card in the photo "has not been shipped" and that the third party only sells a 10 GB unlock. A vendor advert quoted by @analyticeth said the 8 GB modification was "still under development, with major breakthroughs already achieved". Nobody posted a card, a dump or a buyer. In the order posted:

• Earlier claims of the same kind: on 2026-08-24 ti claimed that every 96 GB card on the Chinese market is an unlocked 8 GB card, that only a small number succeed and that the 80 GB unlock is done card by card; no evidence was posted. On 2026-09-18 Arcanum found a Xianyu listing claiming an 80 GB unlock for the 10 GB card "with 60-70g usable"; insomniac pointed out a red square suggesting the screenshot had been edited. The Chinese full-unlock screenshots and remote tests are in the SM, PCIe and Firmware sections below. Source: @jovial_dolphin_29796, 2026-08-24
  Source: @_arcanum_, 2026-09-18
  Source: @ins_mniac, 2026-09-18
  Source: @giaki3003, 2026-09-18
• The informant says the marked cards are 8 GB cards. Source: @ob88594, 2026-10-09
  Source: @ob88594, 2026-10-09
  Source: @ob88594, 2026-10-09
• In theory an 8 GB card could go to 96 GB. Source: @wakeupmbrc1, 2026-10-09
• bigfoot: that would need every memory stack to be non-defective. Source: @bigfoot1144, 2026-10-09
• @djpandaman did not know how to read the 8 GB memory and Gen3/4 leaks on Chinese sites and Twitter: an elaborate pump-and-dump with an easy fake script, or merit. Playing devil's advocate, he said NVIDIA has been sloppy with signing and a production key leak is not impossible, and admitted he was coping. Source: @djpandaman, 2026-10-09
• The vendor advert's line on the 8 GB card. Source: @analyticeth, 2026-10-09
• Without a signed VBIOS @globalapathy does not see how it would be possible. Source: @globalapathy, 2026-10-11
• @djpandaman: in theory 96 GB on some cards and 80 GB on most; the maximum confirmed on a production 8 GB card is 64 GB. Source: @djpandaman, 2026-10-12

What would settle it. An 8 GB card (10de:20c2) showing more than 65024 MiB in nvidia-smi and passing a dense unique-data write and read-back over all of it, with the method posted.
Board. D-072 and D-056 (Dev and research), T-029 (CMP testing) on the board.
More: 8 GB card#Memory.

Contested 8 GB

A screenshot showing 130104 MB (192 GB) on one card

A screenshot in the "Unlock 80GB stable ram on 10G version" feature-request channel appeared to show 130104 MB of VRAM, which some read as 192 GB on one card. In the order posted:

• "i thought it was 192g". Source: @squishysquare, 2026-10-15
• The per-stack arithmetic (25.6 or 21.3 GB per stack) makes no sense. Source: @squishysquare, 2026-10-15
• "VRAM 130104 MB= 65024 MiB x2 = 2x 64GB unlocked 8GB model": two unlocked 8 GB cards reported together. Source: @kio063218, 2026-10-16
• "Fake pic." Source: @reggie7756, 2026-10-19

What would settle it. A single-GPU nvidia-smi -q from the card in question. For comparison, an unlocked 8 GB card reports 65024 MiB.
Board. T-032 (CMP testing) on the board.
More: Memory unlock on the 10 GB card#80 / 96 GB: not real on these dies.

Contested 8 GB and 10 GB

80 GB working on the 10 GB card

In the order posted:

• Rubbra_F said someone got 80 GB on the 10 GB card in the "clean room", a separate Discord server, and linked a message there; ju5t5omedude said similar work had been done there about a month earlier. DaLZy cited a post in the same server for half-enabled stacks and named the clean-room server and Pidan Xiong as two independent confirmations; c133 linked another clean-room message. The clean-room posts cannot be checked from this server's export. Source: @rubbra_f, 2026-08-09
  Source: @ju5t5omedude, 2026-08-09
  Source: @dalzyyy, 2026-08-09
  Source: @c133, 2026-08-09
• @segmond: by the time his card arrived it was clear the 80 GB unlock was not stable, while the 64 GB one was. Source: @segmond, 2026-08-31
• @cuddylac997: when he first joined, a couple of months earlier, he got 76 GB to show up on a 10 GB card and was talked out of continuing; he says he has a warehouse full of cards. Josh questioned his motives; Con3po vouched for him. Source: @cuddylac997, 2026-09-30
  Source: @con3po7016, 2026-09-30
  Source: @_____josh_____, 2026-09-30
• Miséricorde relayed community research that memory above 40 GB "reads and returns meaningful values". Source: @wakeupmbrc1, 2026-10-13
• Some people had cards reading 80 GB, but the memory was unstable. @djpandaman linked a Chinese write-up, "170Hx 10G解锁80G探索" (10G unlock to 80G, an exploration), at blog.kkk.rs. Source: @tekkenz, 2026-10-12
  Source: @djpandaman, 2026-10-12
• @colo2442 gets folding (aliasing) in every experiment, including physical accesses that bypass L2, unless a further unlock frees address bit 14. He asked for the serial number of any card said to work above 40 GB. Source: @colo2442, 2026-10-14
• Miséricorde: the "no folding" statement came from community research, not his own tests. Source: @wakeupmbrc1, 2026-10-14
• @lycorascence recalled papers that got 80 GB working on a 10 GB card at lower memory speed; Miséricorde: "they actually were not"; @scallion_9883: "I believe it is aliasing in L2 now also". Source: @lycorascence, 2026-10-16
  Source: @wakeupmbrc1, 2026-10-16
  Source: @scallion_9883, 2026-10-16
• @ob88594: the "special initialization" the paper mentions is still missing for a full 10 GB unlock. Source: @ob88594, 2026-10-27

What would settle it. A dense unique-data write and read-back above 40 GiB on a named 10 GB card (serial posted), with no folding and no Xid.
Board. T-026, T-027 and T-167 (CMP testing), D-055 and D-131 (Dev and research) on the board.
More: Memory unlock on the 10 GB card#80 / 96 GB: not real on these dies, The 10 GB card: provenance and the 80 GB question.

Contested 10 GB

Videos and listings showing 80 GB on a 10 GB card

A card can be made to report 80 GB even when the memory above 40 GB is not usable, so a capacity readout or a video proves nothing. On 2026-07-29 DeeJay asked about a YouTube video titled "I Unlocked 80GB VRAM on a Mining Card! CMP 170HX + Qwen3.6-27B" (YouTube). Amogh pointed to his earlier message on 10 GB instability; Capone: addressing more than 40 GB crashes; DeeJay noted the video only used about 27 GB, and Studebaker suggested asking the creator to write past 40 GB. barackorama warned that scammers can sell such cards the way fake-capacity SSDs are sold, and decoy said such people had already asked in the server how to unlock. Studebaker: ECC would probably not stabilise the 80 GB unlock. Test any card sold as "80 GB" by writing and verifying unique data above 40 GB before paying the 80 GB price.
Board. D-131 (Dev and research) on the board.
More: Memory unlock on the 10 GB card, Verifying an unlock.

Warning 10 GB
Source: @deejay9983, 2026-07-29
Source: @_j3ns, 2026-07-29
Source: @beefbonecapone, 2026-07-29
Source: @_studebaker_, 2026-07-29
Source: @barackorama, 2026-07-29
Source: @decoy.works, 2026-07-29

A Xianyu seller's "stable 80 GB" on the 10 GB card (July 2026)

In the order posted:

• ti posted photos of a Chinese 170HX said to run stably at 80 GB with an undisclosed method, and noted that a display of exactly 64.000 or 80.000 GiB is possible. Source: @jovial_dolphin_29796, 2026-07-28
• sssyyy: the picture looks fake. Source: @sssyyy4977, 2026-07-28
• Huzifei talked to the seller, who said he runs it at 72 GB and cannot guarantee stability, and offered to teach the method for 50 RMB. Source: @huzifei, 2026-07-28
• adih linked github.com/zerobitcopy/cmp-easyunlock, which he had seen in an "aifond" video, as a possible source. Source: @adih1266, 2026-07-28
• v1ndim said the seller unlocked his card to 80 GB and to "follow Jon Pry's approach"; no model run above 64 GB and no clean gpu-burn result was posted. Source: @v1ndim, 2026-07-28
• Studebaker: assume nothing new until proven, especially from someone selling overpriced 10 GB cards into the hype. Far called it a scam. Source: @_studebaker_, 2026-07-28
  Source: @farawaythrower, 2026-07-28

What would settle it. The same as for "80 GB working on the 10 GB card" above.
More: Memory unlock on the 10 GB card.

Contested 10 GB

A cmpunlocker fork that shows 80 GB (kinako404)

On 2026-08-10 LiYi posted the GitHub fork kinako404/cmpunlocker with a screenshot showing 80 GB and asked whether it is real. giaki3003 replied without text. Nobody confirmed usable 80 GB from this fork.
What would settle it. A dense unique-data write and read-back above 40 GiB on a 10 GB card running that fork.
More: Cmpunlocker forks, Memory unlock on the 10 GB card.

Reported 10 GB
Source: repo github.com/kinako404/cmpunlocker by kinako404 on GitHub, posted by @liyi_00885, 2026-08-10
Source: @giaki3003, 2026-08-10

Sellers claiming a 60 GB unlock on the 10 GB card

On 2026-09-22 Tokukarin's seller claimed that 10 GB cards had been unlocked to 60 GB, and Arcanum said a Goofish seller was holding out on him citing a rumoured 60 GB unlock. Lycorascence: no public unlock goes beyond 40 GB on these cards; Arcanum questioned how it could work physically. When pricing a 10 GB card, treat a seller's 60 GB claim as unverified.
What would settle it. A 10 GB card passing a dense write and read-back above 48 GiB.
More: Memory unlock on the 10 GB card, Buying.

Warning 10 GB
Source: @tokukarin, 2026-09-22
Source: @lycorascence, 2026-09-22
Source: @_arcanum_, 2026-09-22

48 GB on most 10 GB cards

In the order posted:

• @cuddylac997 posted ten 10 GB cards in one host, "ready to run 48gb on each one of these!", and said a 48 GB mode had not been released "but it can be figured out". Source: @cuddylac997, 2026-10-06
  Source: @cuddylac997, 2026-10-07
• @djpandaman: "48 max on some. most cards 40. looking like around 20%". Source: @djpandaman, 2026-10-12
• Miséricorde listed "48g on some (most cards)" among what the 10 GB card can do. Source: @wakeupmbrc1, 2026-10-13
• @djpandaman: 6 of 30 cards he saw tested could do 48 GB. Source: @djpandaman, 2026-10-17

What would settle it. gpcprobe's FBP disable and defective readout plus the unlock result for each 10 GB card owners have.
Board. T-004 (CMP testing), D-027 (Dev and research) on the board.
More: Memory unlock on the 10 GB card#48 GB (the 6th stack, if it is good).

Contested 10 GB

Some 8 GB cards only unlock to 12 or 32 GB

In the order posted:

• chris.via.egnatia quoted the Red Panda Mining store listing for the 170HX: "Unlockable between 12GB to 64GB VRAM! (Silicon Lottery)"; arnold civardagezen could not find that wording. quark2world had not heard of anyone having problems with the 64 GB unlock, though issues could show up later. chris.via.egnatia said a "burnt" mining card would not work at 64 GB and called a purchase "a multilevel gamble"; quark2world and machinezero noted the cards ran in dusty mining farms with no guaranteed lifespan. Source: @chris.via.egnatia, 2026-07-23
  Source: @a.civardagezen, 2026-07-23
  Source: @quark2world, 2026-07-23
  Source: @roger_dammit, 2026-07-23
  Source: @machinezer0, 2026-07-23
• arnold civardagezen noted a wccftech article saying 64 GB is not stable, and jimtjames cited a Chinese blog post saying 64 GB is unstable but 32 GB is stable, wondering whether some early 8 GB cards can only run 32 GB. All four of jimtjames's own cards pass several hours of Vulkan memtest at 64 GB on the 300 W VBIOS. CUTE considered trying a 32 GB-only unlock on his failing card, but none was known to exist. Source: @jimtjames, 2026-08-07
  Source: @a.civardagezen, 2026-08-07
  Source: @cute_sc2, 2026-08-07
• Pti-Tza said Chinese websites claim the Hynix version may run unstably above 32 GB and produce token errors. ju5t5omedude: it sounds like an excuse made up to dump cards that fail the 64 GB unlock; philow expected sellers to dump failed cards. Source: @pti_tza, 2026-08-17
  Source: @ju5t5omedude, 2026-08-17
  Source: @.philow, 2026-08-17

Nobody posted data either way. These are not the genuine 32 GB cards, which are a different board: see 32 GB cards (subsystem 10de:1586).
What would settle it. An 8 GB card that passes a CUDA memory test at 32 GB and fails at 64 GB, with its board number and memory vendor.
Board. T-253 (CMP testing) on the board.
More: 8 GB card#Memory, Testing and burn-in.

Contested 8 GB

Unlocked 8 GB memory is failure-prone

In the order posted:

• "No amount of software can make up for the fact that this was lower binned and prone to failure memory that nvidia had to fuse off." Source: @_____josh_____, 2026-10-05
• @jjb01957: his card's VRAM runs fine and, apart from speed, no different from his 5090. Source: @jjb01957, 2026-10-05
• Miséricorde: "every card runs on ndiv=68 as minimum". Source: @wakeupmbrc1, 2026-10-05

What would settle it. Failure statistics: cards, hours in service, memtest error counts.
Board. T-030 and T-115 (CMP testing) on the board.
More: 8 GB card#Memory, Card failure rates and longevity.

Contested 8 GB

Some cards have only two HBM stacks

On 2026-08-08 a new buyer, 42, had heard that some 170HX cards shipped with only 2 HBM2 stacks and so cannot be upgraded or unlocked. DaLZy dismissed it as Reddit misinformation from July that was still circulating, and nobody in the thread reported such a card. The 32 GB-only cards are a separate case: see 32 GB cards (subsystem 10de:1586).
What would settle it. A memory-layout readout (gpcprobe) or a photo of a card with two stacks.
More: Which card do I have.

Contested 8 GB and 10 GB
Source: @vqo_, 2026-08-08
Source: @dalzyyy, 2026-08-08

The "defective" HBM stacks are good memory

In the order posted:

• A theory @n2dcrypt relayed from another chat: the same stacks are always marked defective because the HBM package height was wrong and they run hotter, so they may not be faulty memory. @globalapathy: at lower voltage and under water they could work, just not optimally. Source: @n2dcrypt, 2026-10-09
  Source: @globalapathy, 2026-10-09
• Petri Krohn: "Almost every 8GB card has FBP 6 and FBP 11 marked as *defective*. This cannot be true!" He reads it as two stages of disabling. Source: @petrikrohn_26497, 2026-10-10
• @djpandaman: in his experience a part marked defective is truly defective, while a disabled part can be enabled. Source: @djpandaman, 2026-10-17
• @lxtend on 10 GB stacks: "they are mostly good even its labeled as defective". Source: @lxtend, 2026-10-20

What would settle it. One marked-defective stack enabled and memtested, at reduced clocks if needed. On the 8 GB card nobody has a way to enable it.
Board. T-029 and T-133 (CMP testing), D-056 (Dev and research) on the board.
More: 8 GB card#Memory, 10 GB card#Memory.

Contested 8 GB and 10 GB

Most cards reach 1.9 to 2.0 TB/s with HBM tuning

@zorg33 believes most cards can reach 1.9 to 2.0 TB/s with HBM fine tuning (he wrote "TH/s"), which he says would scale TP4 decode to 550-600 tok/s. Miséricorde: a 10 GB card with all six stacks would have a 6144-bit bus and "can run 2.5tbps+". No tuned bandwidth figure at that level has been posted. The highest posted so far: 1,741 GB/s on an unlocked 10 GB card at stock, and the NDIV results on 8 GB card#Memory.
What would settle it. HBM bandwidth and gpu-burn correctness at NDIV 70 and 73 on several cards; bandwidth on a 48 GB 10 GB card.
Board. T-143 and T-028 (CMP testing) on the board.
More: Undervolting and power limits#HBM clock (NDIV) and refresh.

Theory 8 GB and 10 GB
Source: @zorg33, 2026-10-06
Source: @wakeupmbrc1, 2026-10-07
Source: @cuddylac997, 2026-10-07

"ECC Enabled" after v0.5 means errors are corrected

In the order posted:

• @zombielurker: nvidia-smi says ECC is enabled, even on his RTX 3090s, but it corrected none of about 5,000 single-bit errors one card throws in memtest_vulkan; gpu-burn behaved the same. Source: @zombielurker, 2026-10-05
• Lab notes: the "ECC Enabled" readout comes from cmpunlocker's reporting patch; no corrected-error counter has been shown moving. Source: @darkmaniac7, 2026-10-21 (lab notes)
• @cute_sc2, arguing the 170HX is no different from an A30 or A100 for serving: "We even have ecc now." Source: @cute_sc2, 2026-10-23

What would settle it. One corrected DRAM error shown on a counter that is not set by the reporting patch.
Board. T-001 (CMP testing), D-032 (Dev and research) on the board.
More: ECC#Does it correct errors? (contested).

Contested 8 GB and 10 GB

SM and clocks

The 8 GB card can go past 74 SMs

In the order posted:

• Miséricorde: every RIR entry except the last reads all ones on the 8 GB card, so it cannot be overridden "without signed falcon code". Source: @wakeupmbrc1, 2026-10-07
  Source: @wakeupmbrc1, 2026-10-07
• @sojun80: "thats just software looking at it and setting a flag in init". Source: @sojun80, 2026-10-07
• @matstatman installed an 8 GB card with 0 defective GPCs that passed the standard 5-minute test. Miséricorde: "if 0 defective then 128"; matstatman: "sooner or later, we will be able to enable them". Source: @matstatman, 2026-10-09
  Source: @wakeupmbrc1, 2026-10-09
  Source: @matstatman, 2026-10-09
• @mountains1605: the path is clear for 10 GB cards, not for 8 GB cards. Source: @mountains1605, 2026-10-10
• @replacedriver47: fuses are read and passed on by software, so it is software-unlockable in theory, "but it would require bypassing security mechanisms". Source: @replacedriver47, 2026-10-27

What would settle it. An 8 GB card (10de:20c2) running more than 74 SMs, with the method posted.
Board. D-072 and D-084 (Dev and research) on the board.
More: Full SM unlock#Why the 8 GB card cannot go past 74, 8 GB card#Limits, and CUTE's 80 SM claim below.

Contested 8 GB

An 80 SM unlock on 8 GB cards (CUTE, unpublished)

On 2026-09-16 CUTE said an AI agent "did something I dont understand" that made all SMs available with the one-time-programmable fuses still in place. On 2026-10-02 he listed unpublished results from agent-driven hacking, including an 80 SM unlock on 8 GB cards; asked whether 80 SMs was really possible, he said yes, but not on every card, and that he would publish later. replacedriver47 had discussed the 80 SM path with Amogh that day and had not been sure it was possible. No method, readout or numbers were posted; 74 SMs remains the public ceiling for the 8 GB card.
What would settle it. An 8 GB card showing 80 SMs in a readout and a benchmark.
Board. T-179 (CMP testing) on the board.
More: Full SM unlock, 8 GB card#Limits.

Reported 8 GB
Source: @cute_sc2, 2026-09-16
Source: @cute_sc2, 2026-10-02
Source: @replacedriver47, 2026-10-02

AI-generated "SEC2 patch" diffs that claim 30 SMs per GPU

On 2026-09-25 arabus posted driver diffs, with comments in Hungarian, for three source files and claimed "FUSE ENABLED 18 SM, SEC2 DISABLED +12=30 SM per GPU". AKuchen rebutted them the next day: those files do not exist in the 610.43.03 driver tree, the log lines the diffs rely on are not in the driver, and the state change they read as a rollback is the enable. Do not apply these patches to a card.
Board. T-180 (CMP testing) on the board.
More: Full SM unlock.

Warning 8 GB and 10 GB
Source: @arabus_74064, 2026-09-25
Source: @akuchen_66498, 2026-09-26

120 SMs and 237 TH/s at 250 W (JetSki)

Lab notes of 2026-10-26 relay a claim credited to JetSki: an SM unlock (patches to d3dx9/cmpunlocker on nvidia-open 580.x) yielding 120 SMs and 237 TH/s at 250 W. JetSki is not a member who posted in the Discord export. The only mention there is JMA1976 on 2026-09-24, posting the install line for the cmpunlocker SMs branch (70 to 74 SMs) and adding that he "heard about it trough Jetski on another mining forum". 120 SMs does not fit an 8 GB card, whose public ceiling is 74 SMs (the lab's unlocked 8 GB cards report 74 on driver 615.71.09). It does fit a 10 GB card after the RIR fuse burn: an unlocked 10 GB card has been reported at about 250 TH/s at 250 W. Which card JetSki ran is not known.
What would settle it. The card variant and patches behind the figure, then a run on that variant.
Board. D-036 (Dev and research) on the board.
More: Full SM unlock, RIR fuse burn.

Theory 8 GB and 10 GB
Source: @jma1753, 2026-09-24
Source: @lxtend, 2026-10-21
Source: @darkmaniac7, 2026-10-26 (lab notes)

Chinese "full unlock" screenshots: 124 SMs, 7936 CUDA cores, 400 W, PCIe 4.0

In the order posted (2026-08-11 unless dated):

• Jack reposted screenshots from a Chinese QQ group, attributed to 皮蛋熊 (Pidanxiong), claiming a full 170HX unlock with 124 SMs, 400 W and PCIe 4.0, and flagged that their authenticity was unknown. Source: @ludashi666, 2026-08-11
• LetterC67: 124 SMs is more than a real A100 has (108). Source: @letterc67, 2026-08-11
• GMZombie007: the device and subsystem IDs shown are not a 170HX's unless the VBIOS was modded. Source: @gmzombie, 2026-08-11
• Humvee55: without the ROM posted, assume it is fake. Source: @scallion_9883, 2026-08-11
• Apoloniato relayed from the same QQ group that a long-time GPGPU trader sells "unlocked" 170HX cards showing 124 SMs and 7936 CUDA cores, more than an A100, at US$2.3k per card including tax and warranty, with remote verification offered for 100 yuan. The trader said he could also unlock 80 GB but it was not worth the time. The group's consensus was that he got something from an NVIDIA internal channel, possibly fuse tricks. Source: @apoloniato_07709, 2026-08-11
• matstatman suggested asking a suspected scammer for something impossible, such as HDMI support. Source: @matstatman, 2026-08-11
• CUTE judged it 90% fake: the screenshots looked like a GRID A100, and 124 SMs would need every normally fused-off unit to work. Source: @cute_sc2, 2026-08-11
• gym: a GRID A100 has 108 SMs, so a cross-flash alone does not explain 124. Source: @jimtjames, 2026-08-11
• Humvee55 offered to confirm or deny it with register reads; otherwise treat it as bogus. Source: @scallion_9883, 2026-08-11
• 2026-09-29: fafinet dismissed a Facebook "80 GB + PCIe 5" claim: the GitHub says 40 GB and PCIe Gen 2. Source: @fafinet, 2026-09-29

What would settle it. A gpcprobe readout and a tensor-core benchmark taken live on such a card by someone other than the seller.
Board. D-065 (Dev and research) on the board.
More: the A100-relative entries under Firmware and VBIOS below, Verifying an unlock.

Contested 8 GB and 10 GB

The 126-SM spec screenshot and Goofish proof-of-concept video

On 2026-09-29 Newman posted a spec screenshot of a modified card and a Goofish listing with a proof-of-concept video (Goofish). mrpandagod doubted it: 126 SMs would mean a die with only one defective TPC, FP16 is shown at 3.5 times the FP32 rate ("how is FP16 at 3.5 * FP32 rate?"), there is no FP16 tensor figure, the SM clock reads 2100 MHz, and a custom script is called for no reason, which suggests market pumping. ti said he had said long ago that someone in China had fully unlocked the 170HX. The published SM unlock for the 10 GB card is on Full SM unlock.
What would settle it. The card itself, with a gpcprobe readout and a tensor-core benchmark.
Board. D-065 (Dev and research) on the board.

Contested 10 GB
Source: @newman_08462, 2026-09-29
Source: @djpandaman, 2026-09-29
Source: @jovial_dolphin_29796, 2026-09-29

OpenGL on unlocked cards ("opengl with ecc")

Asked whether OpenGL is actually enabled on these cards, @matstatman said his coding agent had offered to enable it that day: "opengl with ecc, huang level". Nothing was shown afterwards.
What would settle it. glxinfo output or an OpenGL program running on an unlocked 170HX, with driver and patch named.
Board. T-021 (CMP testing), D-087 (Dev and research) on the board.
More: Other CMP cards and other NVIDIA GPUs (graphics on a 170HX).

Theory 8 GB and 10 GB
Source: @phoenixone7, 2026-10-05
Source: @matstatman, 2026-10-05
Source: @matstatman, 2026-10-05

RT cores unlocked ("It's done.")

The only message in the feature-request channel "Unlock RT cores for CMP cards that have them" is @emogop's "It's done." with no details, link or proof. It may mean the unlock was done or that the request was closed.
What would settle it. emogop's explanation, or a ray-tracing workload with numbers on a CMP card.
Board. T-022 (CMP testing) on the board.

Theory 8 GB and 10 GB
Source: @emogop, 2026-10-25

PCIe and P2P

PCIe Gen4 x16 and "A100 PRO" screenshots

Screenshots collected in cmpunlocker issue #9 show (a) a Chinese marketplace listing for an in-person "170hx to A100 PRO" hard-mod service in Chengdu, three cards minimum, 10 GB cards first, with a card reported as an A100-PCIE-40GB with 126 SMs, 40960 MiB, PCIe Gen4 x16 and ECC on; and (b) a tweet showing a card with 65536 MiB at PCIe Gen4 x16 and 300 W. Neither names a method, and no second person has seen such a card. In the order posted:

• @matstatman posted the tweet ("Oh no, more rumors"), then screenshots of screenshots and a further comment on issue #9. Source: @matstatman, 2026-10-10
  Source: @matstatman, 2026-10-10
  Source: @matstatman, 2026-10-10
• @matstatman: "I think it is a case of old screenhots being mixed". Source: @matstatman, 2026-10-10
• @port00: "Looks like 8GB version"; @dottoreargus: "and an account created today". Source: @port00, 2026-10-10
  Source: @dottoreargus, 2026-10-10
• Lab notes on issue #9: the listing renames the card's device ID, which the public tools do not do; the issue is still open. Source: @darkmaniac7, 2026-10-18 (lab notes)

The Chengdu service may be the same as the paid 10 GB unlock below; nothing confirms it.
What would settle it. lspci -vv showing LnkSta: Speed 16GT/s, Width x16 on both ends for a card the poster owns, with the method named.
Board. D-065 (Dev and research), T-005 (CMP testing) on the board.

Contested 8 GB and 10 GB

PCIe 4.0, ECC, NVLink and twice the SMs on the 8 GB card

In the order posted:

• @endsjeremy: PCIe 4.0 was "pretty definitively off the table without some way to sign the driver". Source: @endsjeremy, 2026-10-11
• @ob88594: there are two separate "8G cracked" reports, a team in Chengdu and someone in this channel. Source: @ob88594, 2026-10-11
• @douya4305: "8g card can also unlock pcie4.0, ecc,nvlink" and "2X sm is on the road"; asked, "right ,just not public". Source: @douya4305, 2026-10-13
  Source: @douya4305, 2026-10-13
  Source: @douya4305, 2026-10-13
• Miséricorde: "no", "at least not public", "thats all really rumored". Source: @wakeupmbrc1, 2026-10-13
  Source: @wakeupmbrc1, 2026-10-13
• @lia_oc: ECC is done in v0.5; the others are not public. Source: @lia_oc, 2026-10-13
• @tarxlazer: "The 8g card of cmp170 unlocks the full core and pcie4. I hope it will come soon." Source: @tarxlazer, 2026-10-14

What would settle it. An 8 GB card with an 8 or 16 GT/s LnkSta, an active NVLink in nvidia-smi nvlink -s, or more than 74 SMs.
Board. D-072 (Dev and research), T-011 (CMP testing) on the board.
More: 8 GB card#Limits.

Contested 8 GB

PCIe Gen3 on an 8 GB card, release promised for 2026-12-09

@rrickk420 (Rickcal) said he sees three paths ("RIR, Samsung memory & falcon") and "I already have gen 3 8GB". He would not share it yet because he does not want to fry people's cards, and said he would release it on 2026-12-09 if nothing else came out by then. The next day he said only some cards can do single-fuse Gen3. The same week @djpandaman mentioned that someone said they had Gen3 on an 8 GB card. No numbers, LnkSta or method were posted.
What would settle it. The release, or lspci -vv showing 8 GT/s on an 8 GB card.
Board. D-059 (Dev and research) on the board.
More: 8 GB card#Limits (what the fuse readouts show).

Theory 8 GB
Source: @rrickk420, 2026-10-09
Source: @rrickk420, 2026-10-09
Source: @rrickk420, 2026-10-09
Source: @djpandaman, 2026-10-09
Source: @rrickk420, 2026-10-10

PCIe Gen4 on the 10 GB card

@arabus_74064 posted a screenshot of a 170HX at PCIe x16 4.0; @bigptoughneigh said it was a 10 GB card with "All the RiR" done. @djpandaman said he has Gen3 and "validated gen 4 on my plx" (his motherboard is Gen3). @cute_sc2 answered yes when asked whether the 10 GB card can use PCIe 4. Gen3 on the 10 GB card is published and reproduced; Gen4 is not: no 16 GT/s LnkSta and no recipe has been posted. @lxtend has Gen3 confirmed, says Gen4 needs "some tricky way to retrain", and has no Gen4 slot to test it.
What would settle it. lspci -vv showing 16 GT/s x16 on a cap-modded 10 GB card in a Gen4 slot, plus a bandwidth test.
Board. T-005 and T-118 (CMP testing), D-028 (Dev and research) on the board.
More: 10 GB card#PCIe, PCIe Gen3 on the 10 GB card.

Theory 10 GB
Source: @arabus_74064, 2026-10-05
Source: @bigptoughneigh, 2026-10-06
Source: @djpandaman, 2026-10-09
Source: @lxtend, 2026-10-15
Source: @cute_sc2, 2026-10-15

Someone has PCIe 4.0 x16 or NVLink working but will not publish (August and September 2026)

In the order posted:

• HumanDrone8721 had been reverse engineering with a QEMU and Ghidra workbench and thought he was close to PCIe 4.0 before his card order was cancelled and he stopped. Source: @humandrone8721_60165, 2026-08-29
• ST: people were working on PCIe 4.0 x16 and one party claimed an NVLink unlock, none of it public; asked for proof, he said there were only claims. On 2026-09-01 he said someone had found how to unlock PCIe 4.0 but feared legal trouble. OB88594 took part in the thread. Source: @stcord, 2026-08-31
  Source: @ob88594, 2026-08-31
• MaitOPS: he is not working on PCIe 4, and nothing he is working on works yet. Source: @_maitops, 2026-09-01

What would settle it. lspci -vv showing 16 GT/s on a 170HX, or active links in nvidia-smi nvlink -s.
Board. T-005 and T-011 (CMP testing) on the board.

Contested 8 GB and 10 GB

Chinese PCIe 3.0 and 4.0 unlocks, and cards with PCIe 4 "out of the box"

In the order posted:

• ST asked whether anyone owns one of the 170HX cards said to run PCIe 4 out of the box; irythros: as far as he knows there has been no proof they exist. Source: @stcord, 2026-09-01
  Source: @irythros, 2026-09-01
• Asked whether the PCIe 3.0/4.0 and NVLink work had progressed, CUTE said there was no new development though everyone was working on it, and that Chinese sources claiming PCIe 3.0/4.0 unlocks were a scam ("not real"), possibly one reason prices rose so fast. Morrowmake: a real PCIe 4.0 unlock would double the cards' value. Source: @a.civardagezen, 2026-09-03
  Source: @cute_sc2, 2026-09-03
  Source: @voidgiraffe, 2026-09-03
• okrasit.: "the chinese unlock to 80gb"; CUTE had yet to see a card doing it. Miséricorde noted Goofish sellers advertising 8 GB unlocks. mrpandagod: anyone with a real internal fuse-writing tool would make the card enumerate as a native A100 40GB. insomniac: reports of stock-unlocked Chinese variants are not conclusively confirmed. Source: @okrasit., 2026-10-01
  Source: @cute_sc2, 2026-10-01
  Source: @wakeupmbrc1, 2026-10-01
  Source: @djpandaman, 2026-10-01
  Source: @ins_mniac, 2026-10-01

What would settle it. A card bought from such a source showing 8 or 16 GT/s in lspci -vv.
Board. T-005 (CMP testing), D-065 (Dev and research) on the board.

Contested 8 GB and 10 GB

A Chinese seller's PCIe 4 and ECC unlock with device ID 10de:20be

In the order posted:

• Null Yang relayed a Chinese seller claiming PCIe 4 and ECC are unlocked on 64 GB 170HX cards, and that "the final (64GB) unlock version will be devid=10de-20be, but not devid=10de-20bd" for mainland regulatory reasons. He inferred the seller flashes an A100 VBIOS rather than using the driver approach, since that approach needs no device-ID change. Source: @alexyzhov, 2026-08-14
• meatsus: possible only with parts of the VBIOS re-signed with NVIDIA's private key, which the community does not have, and the GPU refuses a VBIOS whose ID does not match. Source: @meatsus, 2026-08-14
• CUTE floated buying one and dumping its VBIOS. Source: @cute_sc2, 2026-08-15
• SealClubber: the seller claims the card shows as an A100 GRID, not an A800. Source: @sealclubber, 2026-08-15

Petri Krohn's device-ID theory is under Firmware and VBIOS below.
What would settle it. lspci -nn and a VBIOS dump from a card that went through the seller.
Board. D-065 (Dev and research) on the board.

Contested 8 GB

A Chinese eBay listing: ECC, PCIe 4.0 and 400 W on unlocked cards

In the order posted (2026-08-10):

• Jack reposted screenshots of a Chinese eBay seller claiming a further crack with ECC and PCIe 4.0 on top of the 40 GB and 64 GB unlocks, with a 400 W power limit. Source: @ludashi666, 2026-08-10
• Amogh thinks it is a hoax: 400 W exceeds what the EPS connector is designed for, and only SXM models go to 400 W. He added that they may have signed a VBIOS using the same theory he had. Source: @_j3ns, 2026-08-10
• DaLZy: "not supported and not possible are two things"; a connector could be soldered on. Source: @dalzyyy, 2026-08-10
• weicj relayed the same claim to Amogh. Source: @makesier, 2026-08-10
• gym: they may have ported parts of an SXM VBIOS. Source: @jimtjames, 2026-08-10

What would settle it. lspci -vv, nvidia-smi -q and a sustained power reading from a card that went through the seller.

Contested 8 GB and 10 GB

A Xianyu seller's fuse-burn unlock: PCIe 4.0, all CUDA cores, ECC and NVLink

In the order posted (2026-08-24):

• ti relayed claims from a Xianyu seller that 170HX cards had been unlocked to PCIe 4.0 x16 with all CUDA cores enabled, later also ECC and NVLink. The service was first for organisations ordering 100 or more cards, later for individuals who sign a US$500,000 breach-of-contract penalty; the seller says moving the modified VBIOS to another card has no effect. ti later posted remote-test screenshots from an organisation that bought cards, but admitted the PCIe 4.0 link's "actual speed is only around PCIe 3.0 level". Source: @jovial_dolphin_29796, 2026-08-24
• NFA relayed that Amogh thought it might be real through fuse burning, because the device ID had been changed to one no commercial card uses. Source: @never_fly_away, 2026-08-24
• Inf wanted lspci, nvidia-smi topo -m, NVLink status, and memory and P2P tests before believing it. Source: @inf, 2026-08-24
• Studebaker ("Obvious hype campaign") and Danske called it hype or price manipulation. Source: @_studebaker_, 2026-08-24
  Source: @danske1, 2026-08-24
• Waka324: plausible if someone holds signing capability. Source: @cmdrwaka324, 2026-08-24
• Amogh: "I haven't heard of any cards that have been unlocked with PCIe Gen 4 at ALL", and he recalled no conversation with such a buyer. Source: @_j3ns, 2026-08-24

What would settle it. Inf's list, run on a card in a buyer's hands.
Board. T-005 (CMP testing), D-065 (Dev and research) on the board.

Contested 8 GB and 10 GB

QQ-group claims: a PCIe 4 and NVLink unlock and a "96g sample" sold for 63k CNY

On 2026-09-05 Apoloniato, who relays news from the Chinese 170HX community, posted screenshots from a Chinese QQ group: a PCIe 4 and NVLink unlock attributed to 南小鸟 (Nanxiaoniao), including a "96g sample" said to have sold for 63k CNY (about US$9k), an analysis of the 10 GB version by 皮蛋熊 (Pidanxiong), and other attempt reports. GreenDam said one screenshot was his own work, posted in both places, and that the QQ group mostly discusses GPU prices rather than unlock technology. ST asked what the 170HX advances were; nothing concrete was posted. The same group, which Apoloniato named on 2026-08-12 when asked how to join, was the source of the 124-SM seller reports above. All of it is second-hand.
What would settle it. A buyer of the 96 GB sample, or the unlock shown on a card outside the group.

Reported 8 GB and 10 GB
Source: @apoloniato_07709, 2026-08-12
Source: @rogersmith2026, 2026-08-12
Source: @apoloniato_07709, 2026-09-05
Source: @green_dam, 2026-09-05
Source: @stcord, 2026-09-05

d3dx9/cmpunlocker reaches PCIe Gen 4

In the order posted:

• On 2026-07-19 the fork github.com/d3dx9/cmpunlocker, which claims to reach PCIe Gen 4, came up in #general. Amogh said he would test it, then reported that it "didn't do anything" for him ("This is BS - didn't work for me at all"). Source: @lunaroseoc, 2026-07-19
  Source: repo github.com/d3dx9/cmpunlocker by d3dx9 on GitHub, posted by @_j3ns, 2026-07-19
• Amogh's eGPU is limited to Gen 3, so he asked Far to try. Far's board is also Gen 3, and the script gave him a black screen, which he put down to a wrong driver install. Source: @farawaythrower, 2026-07-20

Nobody confirmed any speed gain from the fork.
What would settle it. The fork run on a Gen4-capable host, with lspci -vv before and after.
Board. T-258 (CMP testing) on the board.
More: Cmpunlocker forks.

Contested 8 GB and 10 GB

PCIe 5.0 on the 170HX

In the order posted:

• Studebaker posted a preview.png apparently showing a PCIe Gen 5 link as sarcastic "proof". MaitOPS: GA100 does not support Gen 5. xanthe_: a similar "gen 5" entry appears somewhere in the nvidia-smi -a output on his own system, so such a reading is not evidence of an unlock. Source: @_studebaker_, 2026-08-24
  Source: @_maitops, 2026-08-24
  Source: @xanthefin, 2026-08-24
• The author of the CMP 100-210 unlock repository claimed PCIe 5.0 on the 170HX in a comment on cmpunlocker issue #9; @68x0 pointed to it: "I wouldn't trust any of this AI slop from this person". Source: @68x0, 2026-10-09
• Amogh (@_j3ns), the cmpunlocker author: "Exactly why I banned them!" Source: @_j3ns, 2026-10-09

What would settle it. lspci -vv showing a 32 GT/s link on a 170HX. Nobody has posted one.
More: Other CMP cards and other NVIDIA GPUs#CMP 100HX (100-210), and the mintoriakamoto/cmpunlocker2 entry below.

Contested 8 GB and 10 GB

mintoriakamoto/cmpunlocker2 promises PCIe Gen5 x16 and 80 GB

On 2026-09-10 DeeJay flagged github.com/mintoriakamoto/cmpunlocker2, whose README section "What gets unlocked" claims PCIe Gen5 x16 on the 170HX and, per clee, the full 80 GB on 10 GB cards. DeeJay: Gen5 is not possible. Schaka mocked it by linking a d3dx9/cmpunlocker commit as "PCIe 6.0 support". Amogh offered to post on X that PCIe Gen 5 is not possible on GA100 SKUs "if that's what it takes to prevent prices from going up further". acedogblast agreed it was too good to be true, especially Gen5 on a design with a Gen4 PHY, and Capone called it "100% slop". Treat its claims as false and do not install it expecting either feature. It is a different repository from akumaburn/cmpunlocker2.
More: Cmpunlocker forks.

Warning 8 GB and 10 GB
Source: repo github.com/mintoriakamoto/cmpunlocker2 by mintoriakamoto on GitHub, posted by @deejay9983, 2026-09-10
Source: @c133, 2026-09-10
Source: @schakaqt, 2026-09-10
Source: @_j3ns, 2026-09-10
Source: @acedogblast, 2026-09-11
Source: @beefbonecapone, 2026-09-11

NVLink working on a 170HX

@djpandaman said on 2026-10-12 he was still getting NVLink to work on his 10 GB cards and "it looks like it will". The paid 10 GB unlock was offered with "nvlink for extra" (see below). On 2026-10-23 @bigptoughneigh said NVLink "works fine, but requires a very simple hardware mod (easier than the x16 mod). It's not my finding to share though." No nvidia-smi nvlink output, photo or method has been posted. @lia_oc (2026-10-22): no NVLink yet, "not even for 10GB that's public"; @ob88594 (2026-10-24): "8G still hasn't been unlocked".
What would settle it. nvidia-smi nvlink -s showing active links on a bridged pair and a P2P bandwidth test over the bridge.
Board. D-025 and D-026 (Dev and research), T-011 (CMP testing) on the board.
More: NVLink#Status of the hardware mod.

Theory 8 GB and 10 GB
Source: @djpandaman, 2026-10-12
Source: @lia_oc, 2026-10-22
Source: @bigptoughneigh, 2026-10-23
Source: @ob88594, 2026-10-24

Mailbox P2P at 6.69 GB/s

In the order posted:

• Manttis measured 6.69 GB/s after "the patch" (not named) on a Huananzhi H12D-8D. Source: @manttis0264, 2026-08-25
• satspace's P2P guide repository retracted its own earlier "Mailbox B2" result of 6.69 GB/s as a false positive: a content test found the peer's VRAM unchanged. Lab notes agree: GA100 mailbox P2P configures cleanly and reports success, yet moves zero bytes with no error and no Xid. The working path is static BAR1 P2P, which satspace measured at 5.30 GB/s one-way and 10.27 GB/s bidirectional with content checks passing. Source: repo github.com/satspace-cpu/cmp170hx-linux-p2p by satspace-cpu on GitHub
  Source: @darkmaniac7, 2026-10-08 (lab notes)
• cibernox measured 6.69 GB/s behind a PEX88096 switch. Source: @cibernox, 2026-10-02

Both Discord figures equal the retracted result to the digit, and neither post says the copy was content-verified.
What would settle it. A destination-content test on the same pair: copy a known pattern to the peer and read it back there. A passing cudaDeviceCanAccessPeer proves nothing.
Board. T-017 and T-277 (CMP testing) on the board.
More: P2P.

Contested 8 GB and 10 GB

P2P latency figures that disagree by orders of magnitude

In the order posted:

• Jeremy (Ends): inside a switch about 8 to 15 ns, off the switch about 30 times slower. Source: @endsjeremy, 2026-07-27
  Source: @endsjeremy, 2026-07-31
• Zorg33: a switch has about a fifth of the CPU's latency. Source: @zorg33, 2026-07-28
• bayley_w: a switch hop is mid single-digit microseconds, irrelevant for pipeline parallel ("10 hops out of a 30 ms step"). Source: @bayley_w, 2026-08-12
• Manttis measured about 6 us before and about 3 us after "the patch" on an H12D-8D. Source: @manttis0264, 2026-08-25
• Jeremy: "5 to 15 ms" for P2P through a PLX card, and later "200ms vs 15ms" per hop. Source: @endsjeremy, 2026-09-04
  Source: @endsjeremy, 2026-09-04
  Source: @endsjeremy, 2026-09-11
• cibernox measured 1.74 us for a pair on one switch and 1.75 us for a pair across switches, behind a PEX88096. Source: @cibernox, 2026-10-02
• markshowa: P2P over the CPU root should be about 1.7 us, about 1.4 us on a crossbar. Miséricorde: the hardware call itself is about 1.5 us. Gnasherrr: a PLX hop is about 105 ns, and a Reddit setup's 14 us came from a retimer plus two switches; Jeremy agreed on the 100 ns range. Source: @markshowa, 2026-10-17
  Source: @wakeupmbrc1, 2026-10-17
  Source: @gnasherrr, 2026-10-17
  Source: @endsjeremy, 2026-10-18
• On real cards at Gen2 x16 with content-verified static BAR1 P2P, satspace measured 1.69 to 1.71 us. Source: repo github.com/satspace-cpu/cmp170hx-linux-p2p by satspace-cpu on GitHub

What would settle it. One measurement of traffic inside a switch and leaving it, with the tool and transfer size stated.
Board. T-190 (CMP testing) on the board.
More: P2P, PCIe switch boards (PEX88096, PEX88064, PEX8747, PLX8749).

Contested 8 GB and 10 GB

P2P inside a Proxmox VM

Asked by Ikyo about Gen2 under Proxmox on 2026-10-03, ElectroHobo said the unlocker's "v.4" had fixed it for him. SeriousRising said dual cards with P2P working in a Proxmox VM had been running for a while, set up largely with an AI agent. No configuration or numbers were posted. Gen2 passthrough under Proxmox has several independent reports; P2P inside a VM has only this one.
What would settle it. p2pBandwidthLatencyTest or nvbandwidth inside the VM with a destination-content check, plus the VM configuration.
Board. T-272 (CMP testing), D-078 (Dev and research) on the board.
More: Proxmox Gen2 passthrough, Virtual machines and passthrough.

Reported 8 GB and 10 GB
Source: @ikyo, 2026-10-03
Source: @nathang6866, 2026-10-03
Source: @seriousrising, 2026-10-03

A x32 link and "Gen2 enabled" under ESXi

On 2026-07-31 nightdays reported the 170HX working flawlessly through VMware ESXi passthrough with PCIe Gen2 enabled. His screenshot showed a x32 link width, which poningru explained: the PCIe Gen1 to Gen5 specifications allow x32, though physical x32 slots are very rare (replacedriver47 mentioned an Asus EPYC board). On 2026-09-07 nightdays gave his recipe: a Dell PowerEdge R7515 with the BIOS MMIO limit at 8TB, 64-bit MMIO parameters on the VM, and the unlock run inside the guest; when Ikyo asked for proof he posted screenshots and a log.
What would settle it. lspci -vv for the same card from inside the guest and from the host.
More: BAR1 64 GB on ESXi, Virtual machines and passthrough.

Reported 8 GB
Source: @.nightdays, 2026-07-31
Source: @poningru, 2026-07-31
Source: @replacedriver47, 2026-07-31
Source: @.nightdays, 2026-09-07
Source: @ikyo, 2026-09-09

More than 8 high-VRAM GPUs fail in one server because of BAR limits

robbbbbb7973 relayed on 2026-07-24 that someone on the Discord had trouble running more than 8 high-VRAM GPUs in a single server because of BAR limitations. It is second-hand, with no board or BIOS details. Con3po separately said BAR1 had to be handled at 8 cards because of a limit around 483 GB (see the next entry).
What would settle it. The board, BIOS and MMIO settings of a host that fails at the ninth card.
More: Resizable BAR and BAR1 64 GB, Large rigs (8 or more cards).

Reported 8 GB and 10 GB
Source: @robbbbbb7973, 2026-07-24

Con3po's 8-card switch pod: which uplink carried the 11-hour zero-error run

Con3po runs 8 CMP 170HX at Gen2 x8 each behind a PEX88096 switch driven from a Minisforum S1 Max mini PC, with P2P on the switch so card-to-card traffic never touches the narrow uplink (Zorg33 confirmed P2P runs over the full x16 through a switch). On 2026-08-28 he described the uplink as a PCIe 4.0 x4-to-OCuLink chain that replaced an earlier USB4 dock link, and said BAR1 had to be handled ("mailbox bar1") at 8 cards because of a limit around 483 GB (wording unclear). On 2026-09-07 he reported 11 hours at full load with zero Xid and zero PCIe errors over a single 40 Gb/s USB4 uplink, adding that one USB4 tunnel drop takes all 8 cards down. A 2026-09-05 note says his cheap OCuLink parts never connected. The records disagree on which link the 11-hour run used.
What would settle it. Con3po's answer, or the run repeated on a named uplink.
More: PCIe switch boards (PEX88096, PEX88064, PEX8747, PLX8749), Large rigs (8 or more cards).

Reported 8 GB and 10 GB
Source: @con3po7016, 2026-08-25
Source: @zorg33, 2026-08-25
Source: @con3po7016, 2026-08-28
Source: @con3po7016, 2026-09-07

A 6-port PEX88096 board, and more modes on fixed boards

PEX88096 switch boards present x16 to the host and fan out to 5 cards at x16, 10 at x8 or 20 at x4. On 2026-08-25 robbbbbb7973 suspected that Broadcom firmware could unlock other modes on boards fixed at one split, such as Zorg33's board fixed at 10x8. On 2026-09-29 mrpandagod mentioned a 6-port variant whose layout (possibly not 3x x16 or 6x x8) nobody has confirmed, and Arcanum noted that one popular board exposes only 48 of its 80 downstream lanes. None of this has been tested.
What would settle it. The 6-port board's lane map from its configuration file or a photo of its slots; a fixed board reflashed to another mode.
Board. D-140 (Dev and research) on the board.
More: PCIe switch boards (PEX88096, PEX88064, PEX8747, PLX8749).

Theory 8 GB and 10 GB
Source: @robbbbbb7973, 2026-08-25
Source: @zorg33, 2026-08-25
Source: @djpandaman, 2026-09-29
Source: @_arcanum_, 2026-09-29

Firmware and VBIOS

Paid 10 GB unlock for about 500 USD in Chengdu

Relaying an informant, @ob88594 said a third party sells a paid 10 GB unlock: "$500, get more than A100 40G PCIE, nvlink for extra", in Chengdu. @aaasss1456 said a friend "went to the factory and made a 10g card", that it "displays as an a100" with "8912 cuda core" (a full 128-SM GA100 has 8192), and that it cost nearly 500 USD. @analyticeth quoted the vendor's advert: all changes are permanent inside the card, no unlock software or driver patches are needed, and the card behaves like an original A100. Miséricorde read it as an RIR macro burn. No buyer has posted lspci, nvidia-smi -q or a fuse dump from such a card.
What would settle it. A buyer posting lspci -nn, nvidia-smi -q and a gpcprobe dump of a card that went through the service.
Board. D-065 (Dev and research) on the board.
More: 10 GB card#Buying.

Theory 10 GB
Source: @ob88594, 2026-10-09
Source: @ob88594, 2026-10-09
Source: @aaasss1456, 2026-10-09
Source: @aaasss1456, 2026-10-09
Source: @analyticeth, 2026-10-09
Source: @wakeupmbrc1, 2026-10-09
Source: @aaasss1456, 2026-10-10
Source: @aaasss1456, 2026-10-10

Unlocks sold as a service, with cards as payment

@ob88594's account of the Chinese market: merchants pay unlockers in cards ("The fewer you break, the higher your reward"), then sell unlocking as a service. The unlock merchants on Xianyu took their service down after someone bought out the capacity, and the 10 GB unlock had circulated since early August among merchants who bought the information, which he says is why unknown buyers took the 10 GB cards. Only his informant photographed an "80G version", and the unlocker gave no details. For the 8 GB card there was "no public information" from the Chinese side.
What would settle it. A listing, receipt or buyer report for the service.
Board. D-065 (Dev and research) on the board.

Theory 10 GB
Source: @ob88594, 2026-10-12
Source: @ob88594, 2026-10-12
Source: @ob88594, 2026-10-12
Source: @ob88594, 2026-10-12
Source: @ob88594, 2026-10-12

20 donor cards, 5 bricked during the 10 GB work

In the order posted:

• @ob88594, marked as hearsay: private groups get free hardware from merchants in exchange for early information; "the 10G unlock came about because a merchant gave them 20 cards; 5 were bricked during testing". Source: @ob88594, 2026-10-09
• Miséricorde relayed it in #hardware and guessed they tried enabling defective blocks or stacks. Source: @wakeupmbrc1, 2026-10-09
  Source: @wakeupmbrc1, 2026-10-09
• @djpandaman does not see how a cold reset would not recover a card: his sixth-stack errors that day always cleared on a cold boot. He agreed that doing unlocks one at a time through RIR can get you into trouble. Source: @djpandaman, 2026-10-09
  Source: @djpandaman, 2026-10-09

What would settle it. One of the bricked cards, with what was written to it.
Board. T-024 and T-036 (CMP testing) on the board.
More: Driver and unlock failures#Bricked cards during 10 GB development (hearsay), RIR fuse burn.

Contested 10 GB

A private unlock from "a guy with 400 cards"

On 2026-10-17, while the channel discussed whether anyone has NVIDIA's signing keys, Miséricorde said a private unlock result he had seen came from "guy with 400 cards" who "doesnt sell", and that "either tool is real or someone got the keys". The regulars answered that nobody on the server has them. @djpandaman: "be cautious, it's easier to fake a screenshot when there's money on the table than to do it legitimately." Which card and what result were not posted.
What would settle it. The tool, or its result on a second owner's card.
Board. D-065 (Dev and research) on the board.

Theory 8 GB and 10 GB
Source: @wakeupmbrc1, 2026-10-17
Source: @wakeupmbrc1, 2026-10-17
Source: @wakeupmbrc1, 2026-10-17
Source: @djpandaman, 2026-10-17
Source: @djpandaman, 2026-10-17

An earlier private "dirty" unlock under NDA

@zingy90210 (PJ) said a talk Petri Krohn gave covered NVIDIA's firmware signing, that "Petri, Jon and 3 others managed to do 'dirty' work and unlock it. Think NVIDIA made jon sign an NDA", and that they are now trying to turn their process into a clean unlock. He gave no link to the talk, and nobody corroborated it. Petri Krohn is on the server and can be asked directly.
What would settle it. A link to the talk, or Petri Krohn's own account.
Board. D-065 (Dev and research) on the board.

Theory 8 GB and 10 GB
Source: @zingy90210, 2026-10-18
Source: @zingy90210, 2026-10-18
Source: @zingy90210, 2026-10-18

A local model re-derived the v0.5 patch

In the order posted:

• @exile_1ck: when he asked Claude to install v0.5, it found that his local Qwen 3.8 Flash Next, told to reverse engineer the BIOS with morluto/rea, had produced the same patch the day before. He added that the model also made Strata faster than he had seen it run unmodified. Source: @exile_1ck, 2026-10-07
  Source: @exile_1ck, 2026-10-07
  Source: @exile_1ck, 2026-10-07
• @djpandaman: "did it actually unlock it from scratch or did it just use the ecc branch before it was merged to master branch". eXile: "i am not entirely sure". Source: @djpandaman, 2026-10-07
  Source: @exile_1ck, 2026-10-07

What would settle it. The model's diff and session log, compared with the upstream ECC branch.
Board. D-066 (Dev and research) on the board.

Contested 10 GB

Unlocks have permanent effects

@rrickk420 said on 2026-10-09 that since June his theory has been that someone would "sneak in brick fuse burns" into an unlock. On 2026-10-25 @stcord posted an X post by unixpickle with "Question everything, folks"; @segmond then said he would not unlock anything more until he is sure there are no permanent effects. The post's content is not in the Discord export. For context, the Driver unlock page states that cmpunlocker flashes and burns nothing, while the RIR fuse burn on 10 GB cards is permanent by design.
What would settle it. The post's actual claim, or a fuse dump taken before and after a plain cmpunlocker install and a cold boot.
Board. D-115 (Dev and research) on the board.

Theory 8 GB and 10 GB
Source: @rrickk420, 2026-10-09
Source: @stcord, 2026-10-25
Source: @segmond, 2026-10-25

A Reddit claim that the unlock bricks cards through "poisoned drivers"

In the order posted (2026-07-24):

• A Reddit comment (reddit.com/r/LocalLLaMA/s/7yYZpuHixU) claimed the unlock bricks cards through "poisoned drivers". Rickcal: no bricking had been reported on the server, and the user was almost certainly pushing a 10 GB card to 80 GB, which does not work above 40 GB. Source: @rrickk420, 2026-07-24
• Eksitus read it the same way. Source: @eksitus, 2026-07-24
• Samuel Hayden: the mod makes no permanent changes, it "literally edits a couple registers", and the card reverts to stock. Source: @samuel_hayden., 2026-07-24
• gym showed that the user's mmapeak screenshot was a re-crop of Far's result (identical numbers), also reused by a Chinese blog, so it did not prove the user even owned a card. Source: @jimtjames, 2026-07-24

What would settle it. A bricked card, with the driver and the steps that led to it.
More: Driver unlock, Recovery.

Contested 8 GB and 10 GB

A VBIOS TOCTOU test damaged a card for good

@youngju10000 (영쥬) wrote that he "saw in a post" that a VBIOS TOCTOU test caused permanent damage that restoring the original BIOS did not fix. The post was not linked and has not been found in the export.
What would settle it. The original post.
More: VBIOS and firmware.

Theory 8 GB and 10 GB
Source: @youngju10000, 2026-10-08

An unlocked VBIOS for the 10 GB card

In the order posted:

• @supasnipa360: "hearing its been done for the 10gb variant as well but a harder find it seems due to legal concerns". Source: @supasnipa360, 2026-10-18
• @xanthefin: the unlock makes no VBIOS change in the first place. @globalapathy: someone able to write a VBIOS would not have stopped at 64 GB. Source: @xanthefin, 2026-10-18
  Source: @globalapathy, 2026-10-18

What would settle it. The file, compared with the stock 10 GB VBIOS (92.00.66.00.02).
More: VBIOS and firmware#The "64 GB unlocked VBIOS" repos (bogus).

Contested 10 GB

Why driver-loaded VBIOS overrides work (Petri Krohn's explanation)

Petri Krohn posted "Ampere VBIOS modding explained", an account of why the driver can hand the GPU's microcontrollers modified VBIOS data. It is an explanation, not a measurement, and nobody has checked it against a dump.
What would settle it. A dump that confirms or corrects it.
Board. T-016 (CMP testing), D-060 (Dev and research) on the board.
More: VBIOS and firmware#Background: why driver-loaded overrides work (Petri Krohn).

Theory 8 GB and 10 GB
Source: @petrikrohn_26497, 2026-10-20

A complete 10 GB unlock exists privately

On 2026-10-07 Miséricorde said the 10 GB card "can be unlocked into full ga100": 128 SMs, 48 GB, six-stack bandwidth, "nvlink and 4.0 also"; 40 to 48 GB is "legit", 80 GB a "big maybe". The same day @cuddylac997 posted overnight tests of unlocked 10 GB cards, measured 1,741 GB/s, and said "There's no releasing it" and "not every card unlocks the same or in the same way". Since then the 128-SM unlock and Gen3 on the 10 GB card have been published and reproduced (see Full SM unlock and PCIe Gen3 on the 10 GB card). The NVLink and PCIe 4.0 parts of the claim have not been shown.
What would settle it. See the Gen4 and NVLink entries above.
Board. D-073, D-025 and D-028 (Dev and research) on the board.

Theory 10 GB
Source: @wakeupmbrc1, 2026-10-07
Source: @wakeupmbrc1, 2026-10-07
Source: @wakeupmbrc1, 2026-10-07
Source: @wakeupmbrc1, 2026-10-07
Source: @cuddylac997, 2026-10-07
Source: @cuddylac997, 2026-10-07
Source: @cuddylac997, 2026-10-07

A Russian Telegram group's full unlock "ready for two weeks" (July 2026)

lorfidel relayed posts from a Russian Telegram announcement group (about 4,000 observers, per Amogh) run around a person called "HPC". On 2026-07-17 they claimed the full unlock had been ready for two weeks and that the poster had flown to the developer to test it on his own hardware. The same group was selling cards in lots of 10 at 38,500 rubles each (about US$500), and HPC said the unlock would be released the next day. meatsus was openly sceptical.
What would settle it. Superseded: what HPC released the next day is the next entry.

Reported 8 GB and 10 GB
Source: @lorfidel, 2026-07-17
Source: @_j3ns, 2026-07-17
Source: @meatsus, 2026-07-17

The modified 610-driver unlock released through HPC: clean or not

In the order posted (2026-07-18):

• lorfidel: HPC posted an unlock based on a modified NVIDIA 610-series driver, whose developers are only known to be from China. He argued that personal use carries no risk and cited modified-driver precedents (Shinjo Kurumi, dartraiden's nvidia-patcher). Source: @lorfidel, 2026-07-18
• Thanos was unsure whether it was "clean" ("We got beaten though"). Source: @thanos8912, 2026-07-18
• meatsus urged everyone to stay away and not promote it in the Discord: "its proven that it can be unlocked without anything illegally leaked". Source: @meatsus, 2026-07-18
• Amogh also preferred to stay away, saying he was close to an actually clean unlock. Source: @_j3ns, 2026-07-18

Petri Krohn's account describes the leaked "Chinese unlock" as patched version 610 open-source drivers, which matches this release (next entry). This wiki carries only the clean unlock: Driver unlock.

Contested 8 GB and 10 GB

The "Chinese unlock" was a fake, or the original group's leaked proof of concept

In the order posted:

• aidoc: the "chinese leaked unlock" is a fake; a few Goofish sellers were faking closed-door success on the chip, had made no progress on PCIe speed and were only pushing prices up. Source: @sekkit_75019, 2026-07-22
• Eksitus asked whether it was instead the unlock code stolen from the original team, linking Petri Krohn's account. Source: @eksitus, 2026-07-22
• Humvee55: the "chinese" unlock was word for word the original group's proof of concept, and the stolen proof of concept was all they ever had. Source: @scallion_9883, 2026-07-22
• giaki3003: sellers were marketing the card as an AI revolution hoping for A100 pricing. Source: @giaki3003, 2026-07-22
• ju5t5omedude: the unlock team "could not even" keep the "dirty unlock" from leaking. Source: @ju5t5omedude, 2026-07-27

Petri Krohn's account says the leaked driver's diff matched the original group's driver modification guides word for word. Source: @petrikrohn_26497, 2026-07-22 A longer write-up of the unlock's history by @petrikrohn_26497 exists; see Discord (pointer only, until the author agrees). aidoc's "fake" may refer to the sellers' further PCIe claims rather than to the leaked driver itself.
What would settle it. aidoc saying which claim he meant.
More: Unlock history: who found what, and when.

Contested 8 GB and 10 GB

The unlock circulated in China a month before it went public

Rickcal said on 2026-07-30 that he had heard the unlock was out in China about a month before it became public, so Chinese holders may have unlocked and benchmarked earlier. Arcanum noted that the Bykski 170HX water block sold for about US$100 before the unlock went public and about US$250 by 2026-07-30, and that cards had vanished from Taobao by Friday 2026-07-17, so Chinese sellers already knew. Hearsay with no direct evidence.
What would settle it. A dated Chinese post or benchmark from before mid-July 2026.
More: Unlock history: who found what, and when, Price history and market.

Theory 8 GB and 10 GB
Source: @rrickk420, 2026-07-30
Source: @_arcanum_, 2026-07-30

"gen3" on 2026-08-08 was an AI agent's post, not an unlock

On 2026-08-08 at 17:32 UTC R3L34S3TH3KRCN posted "gen3" twice, and Studebaker cheered that PCIe Gen 3.0 had been unlocked. R3L34S3TH3KRCN then apologised: a ChatGPT agent he had asked to search and summarise progress had posted the research item into the chat. It is not a Gen3 unlock and not a Gen3 date; Humvee55 said that day that Gen3 work was still in progress.
More: PCIe Gen3 on the 10 GB card, Unlock history: who found what, and when.

Reported 8 GB and 10 GB
Source: @r3l34s3th3krcn, 2026-08-08
Source: @_studebaker_, 2026-08-08
Source: @scallion_9883, 2026-08-08

Satire is not a finding

Several posts in August 2026 were deliberate trolling and are not results: a "170hx dear edition has HBM4", "unlocked 128GB on my 170hx" via Claude, "248 sm's and pcie6.0", "Nice/sec" benchmarks and a "wireless unlocked 128gb 90hx" joke (2026-08-17). Members joked about spreading fake screenshots to mislead sellers, and a seller did pick one up (see the "42069 special" entry under Buying). On 2026-08-22 members joked about spreading lines such as "33% faulty memory", "missing VRM stages" and "PCIe gen1" to cool demand: satire, not findings. Arcanum noted seriously that many mining cards lack provenance; irythros: "you dont have to lie about the card to make it seem bad. You just dont need to say the good part".

Reported 8 GB and 10 GB
Source: @giaki3003, 2026-08-17
Source: @_studebaker_, 2026-08-17
Source: @kellertreppe123, 2026-08-17
Source: @irythros, 2026-08-17
Source: @matstatman, 2026-08-17
Source: @sealclubber, 2026-08-17
Source: @_arcanum_, 2026-08-22
Source: @irythros, 2026-08-22
Source: @quark2world, 2026-08-22

The Chinese seller can sign fuse-burn operations

On 2026-08-24 the channel asked how the Chinese seller could sign fuse-burn operations. bayley_w asserted that an NVIDIA signing key had been obtained; AKuchen asked whether that was verified, and it was not. ti guessed lab access to signing, DaLZy had heard of an internal nvflash version under NDA, and ti relayed that the seller says the signing is not tied to his identity. AKuchen: the keys are a trade secret, and "If he uses signed keys we cannot reproduce this for legal reasons." Con3po asked whether signing is per card (bound to the chip ID), so that a per-chip file could be sold without shipping the card; nobody answered. No key material was posted, and none belongs on this wiki.
What would settle it. Nothing this wiki will carry. A card that went through the seller, examined by its buyer, is the most anyone can check.

Theory 8 GB and 10 GB
Source: @bayley_w, 2026-08-24
Source: @akuchen_66498, 2026-08-24
Source: @jovial_dolphin_29796, 2026-08-24
Source: @dalzyyy, 2026-08-24
Source: @con3po7016, 2026-08-24

NVIDIA's dev and production keys are the same

In the order posted:

• PJ claimed that NVIDIA's dev and production keys are the same; insomniac corrected that only the code structure is similar, not the key, otherwise the work would have been unnecessary. Source: @zingy90210, 2026-09-28
  Source: @ins_mniac, 2026-09-28
• The next day PJ repeated the claim. Waka324 corrected it: the keys are different, the source is the same, and "A blown fuse prevents the dev key from being used." Phoenix17 argued that source matters less now that AI makes disassembly easy. Source: @zingy90210, 2026-09-29
  Source: @cmdrwaka324, 2026-09-29
  Source: @phoenixone7, 2026-09-29

Only the fact of the discussion is recorded here; no key material was posted.

Contested 8 GB and 10 GB

Linus Tech Tips is on the server ("LTT video soon")

On 2026-09-09 Rickcal claimed Linus (LTT) was in the Discord and that people should stock up. Adust pointed out that the account had been created the day it joined and was named ltt170hxvideosoon_18811: a joke account, not LTT. Press coverage of the card is listed on Press and coverage.

Reported 8 GB and 10 GB
Source: @rrickk420, 2026-09-09
Source: @adusts, 2026-09-09

Original members moved to another server

Rickcal claimed on 2026-09-17 that many original members had moved to a different server because of FUD on this one. Nothing else in the export confirms it.

Reported 8 GB and 10 GB
Source: @rrickk420, 2026-09-17

A remote test of a Chinese "full unlock": PCIe 4, device 0x20BE, 7936 CUDA cores

In the order posted (2026-08-12):

• weicj: his group tested the seller's machine through a remote shell the seller opened: PCIe 4 "fully unlocked", 18.52 GB/s one-direction PCIe bandwidth (the community Gen2 unlock reached about 6 GB/s then), device ID 0x20BE (GA100), 7936 CUDA cores and 400 W maximum power listed. He conceded that the only thing they could confirm was the device ID, not that the card is a 170HX. Source: @makesier, 2026-08-12
• bayley_w: PCIe 4 is about 32 GB/s per direction; it could be a GRID A100 with the device or product ID intercepted in the kernel. He asked whether tensor FLOPS were measured; there was no answer. Source: @bayley_w, 2026-08-12
• AKuchen: the seller's screenshot shows subsystem ID 0x134F, an A100 SXM4 subsystem, so it looks like a possibly hacked A100-family card (device 0x20BE, 124 SMs), not a 170HX: "There is currently no confirmed hacked 170hx". He asked whether the bandwidth shown matched a 170HX (about 1.5 TB/s) or an A100 80GB (2 TB/s). Source: @akuchen_66498, 2026-08-12
• Petri Krohn: the device ID is fused into the silicon, while the subsystem ID comes from the VBIOS. bayley_w: a 4-stack GA100 "GRID A100" for MIG users was planned but never shipped. Source: @petrikrohn_26497, 2026-08-12
  Source: @bayley_w, 2026-08-12
• weicj: the IDs can be modded, and 7936 CUDA cores exceeds any stock A100. Source: @makesier, 2026-08-12

What would settle it. A tensor-core benchmark and the memory bandwidth on that machine, and the card's board number.
Board. D-065 (Dev and research) on the board.

Contested 8 GB and 10 GB

Chinese "A100 7936SP" variants: IDs and VBIOS versions

Petri Krohn listed the two Chinese "A100 7936SP" versions: 40 GB = 10DE 20BE 10DE 144E with VBIOS 92.00.36.00.05; 96 GB = 10DE 20BF 10DE 144E with VBIOS 92.00.79.00.01. The 96 GB VBIOS is not in the TechPowerUp database, and the 92.00.36.00.05 listed there belongs to an A100 SXM4 80GB or an NVIDIA GRID A100A 32GB (10DE 20B2 10DE 1463), neither of which matches the Chinese card. He suspects the Chinese are cross-flashing VBIOS files. The seller screenshot in the entry above showed subsystem 0x134F, not 144E.
More: Other CMP cards and other NVIDIA GPUs.

Reported
Source: @petrikrohn_26497, 2026-08-12

96 GB A100s exist, and "enhanced" A100 listings in China

Luna (2026-07-20) noted that an OEM-exclusive A100 96 GB exists but is absent from NVIDIA's documents, and on 2026-08-26 said 96 GB A100s exist in PCIe and SXM4 form, roughly the top 1% of A100s ("It's not a rumor, they exist"); larrylegend posted a photo, and Con3po once saw one for about US$3k but found it too sketchy. bayley_w (2026-08-06) listed odd A100s: a 32 GB automotive version with no closed-loop clock control that randomly pulls 400 W, a 64 GB engineering sample and rare 96 GB units. SkirtChaser found a wccftech article on a 96 GB A100 with 7936 cores, and judged a 96 GB A100 eBay listing that Studebaker found suspicious (photos reused across listings). On 2026-08-24 Inf posted Xianyu photos of a card advertised as 96 GB (RMB 62,888). AKuchen: the photos show an A100 80GB SXM4 reporting about 100 GB, not a 170HX; ti: the RMB price is made up, only a way to message the seller privately. MaitOPS linked a Tom's Hardware article (April 2024) on "enhanced" A100s in China with 7,936 CUDA cores and 96 GB HBM2, and suggested the core-count screenshots relate to those rather than to a CMP.
More: Other CMP cards and other NVIDIA GPUs.

Reported
Source: @lunaroseoc, 2026-07-20
Source: @_j3ns, 2026-07-20
Source: @yuiayanami, 2026-07-20
Source: @farawaythrower, 2026-07-20
Source: @skirtchaser, 2026-08-06
Source: @bayley_w, 2026-08-06
Source: @robbbbbb7973, 2026-08-06
Source: @_studebaker_, 2026-08-06
Source: @giaki3003, 2026-08-06
Source: @inf, 2026-08-24
Source: @jovial_dolphin_29796, 2026-08-24
Source: @akuchen_66498, 2026-08-24
Source: @_maitops, 2026-08-24
Source: @_j3ns, 2026-08-26
Source: @lunaroseoc, 2026-08-26
Source: @larrylegend3504, 2026-08-26
Source: @con3po7016, 2026-08-26

Which card the Chinese "full unlock" started from

weicj pointed at TechPowerUp VBIOS 282151 (an "NVIDIA A100" VBIOS) as possibly the base of the Chinese fully unlocked card, and bayley_w identified it as either a 40 GB engineering sample or a PG199 board. weicj said the seller has worked on PG199 for about three years, has bought up most PG199 boards in China and claims an NVLink secret for PG199. On 2026-09-01 MaitOPS suggested that the "Unlocked CMP 96GB 7936SM" card a Chinese seller showed earlier is not a CMP at all but the card in an ithome article: a 96 GB, 7936 SP A100 on Xianyu, which he reads as a GRID A100B engineering sample that accepted an A100 SXM VBIOS. Not verified.

Theory
Source: @makesier, 2026-08-12
Source: @bayley_w, 2026-08-12
Source: @_maitops, 2026-09-01

The Chinese unlock changes the device ID and flashes an A100 VBIOS (Petri Krohn's theory)

Petri Krohn (2026-08-12): device ID 10DE 20BE was never used by a released NVIDIA product and appears only on Chinese 96 GB A100 cards. His theory is that more fuses are burned to change the device ID to one that accepts an A100 VBIOS, that the fuses disabling HBM2 stacks are also overridden, and that the same approach is being applied to the 170HX. On 2026-08-20 he argued that most of the 170HX's limits are in the VBIOS, so with an A100 VBIOS the card would act almost like an A100, and that what stops it booting one is the device ID fused into the die. He concluded the Chinese PCIe 4.0 unlock is a hardware modification (device ID changed to 20BE, the SPI ROM flashed with an A100 VBIOS), citing the vendor statement in the "device ID 10de:20be" entry under PCIe. He also noted an open problem: fuses can only be set, never cleared, so not every target ID is reachable, and how the change works was unexplained. Most members considered burning fuses on a card this expensive too risky to try.
More: RIR fuse burn (fuse burns are permanent), Fuses and floorsweep maps.

Theory 8 GB and 10 GB
Source: @petrikrohn_26497, 2026-08-12
Source: @petrikrohn_26497, 2026-08-20

How the Chinese seller is said to work: VBIOS, fuse and BootROM cracking, NDA sales

weicj said (2026-08-12) his group suspects the seller cracks not only the VBIOS but also fuses and the BootROM; AKuchen said overriding one-time-programmable fuses would need leaked NVIDIA data and predicted it would leak unless hardware modding is involved. Per weicj, the seller sells to organisations under an NDA with a 3-year resale restriction, cites a US$500k penalty if his work leaks, is cautious with individuals, aims to sell fully unlocked 8 GB cards above US$3,000, and was buying back 8 GB and 10 GB 170HX in China at high prices. weicj kept "80% confidence" that the PCIe hack on the 170HX is real, and said NVLink was not claimed for the 170HX. On 2026-08-30 ST said the Chinese modders' VBIOS mods are locked to serial numbers ("added DRM to their bios mods"); asked how they could sign them, he said there is a rumour they have a signing key, and stressed that it is hearsay. No key material was discussed.

Theory 8 GB and 10 GB
Source: @makesier, 2026-08-12
Source: @akuchen_66498, 2026-08-12
Source: @stcord, 2026-08-30
Source: @realfreaky, 2026-08-30

An "A100 SXM4 80GB with PCIe fingers" at about US$9.5k with a DRM'd BIOS

Apoloniato shared screenshots (2026-08-24) of a new product from "the well known Chinese GPU hacker": a card that reports as an A100 SXM4 80GB, has PCIe fingers and a 400 W limit (against 300 W normally), priced at US$9.5k. ST noted it uses a DRM'd BIOS. The maker claims that cards with all six HBM dies working are rare; Apoloniato thought cards with five good stacks should be much more common. B16_c speculated the firmware might have been put onto a 170HX, and ST wondered whether lower memory clocks could revive broken stacks. For comparison, an A100 80G was listed at US$13k in China and an A100 40GB at about US$4k on eBay.

Reported
Source: @apoloniato_07709, 2026-08-24
Source: @cmurrayis, 2026-08-24
Source: @stcord, 2026-08-24
Source: @matstatman, 2026-08-24

Reworked A100s: SXM dies reballed onto PCIe boards

On 2026-08-08 the chat discussed odd A100 cards sold out of China, including a Reddit post about an "A100 SXM4 64GB". gym and quark2world described reworkers who lift GA100 dies off SXM modules and reball them onto stock A100 PCIe boards; quark2world says Greg of gpulab does this, and with a stock cooler fitted the card looks genuine until you run lspci, while gym has seen the same done with H100. quark2world posted a screenshot about NVLink on converted cards, sourced from gpulab rather than his own testing. Check any "A100" you buy with lspci -nn.

Reported
Source: @jimtjames, 2026-08-08
Source: @quark2world, 2026-08-08

A 64 GB A100 in a PCI ID list

quark2world, reverse engineering with an AI agent, found an NVIDIA PCI ID list that appears to include a 64 GB A100 variant ("There was a 64GB A100?! (Or did my clanker hallucinate?)"), and said it seems real after checking. Only a screenshot was shared, with no ID values transcribed. The Reddit "A100 SXM4 64GB" is in the entry above.

Theory
Source: @quark2world, 2026-08-08

PG199 (DRIVE A100) unlocks to 96 GB

In the order posted:

• York Zhang noted that the NVIDIA DRIVE A100 (PG199) has a 6144-bit bus and floated unlocking its memory to 96 GB, citing the TechPowerUp spec page. PG199 supply is scarce. Source: @yorkzhang_37447, 2026-07-20
  Source: @robbbbbb7973, 2026-07-20
  Source: @decoy.works, 2026-07-20
• Amogh shared a PG199 eBay listing at about US$6k, saying it may have a different VBIOS but that is no excuse to spend that much. Humvee55: buy the SXM2 version for a little over US$1k plus an SXM2-to-PCIe daughterboard for about US$50; he had never seen a legitimate PCIe version and thought the SXM2 version could probably go past 32 GB. Source: @_j3ns, 2026-07-28
  Source: @scallion_9883, 2026-07-28
  Source: @cute_sc2, 2026-07-28
• Asked by Waypoint, Amogh said he may include the PG199 in cmpunlocker since it is a GA100, but doubts it works because it could be binned and fused off like other A100s. Source: @waypoint3642, 2026-08-10
  Source: @_j3ns, 2026-08-10
• Wayz ran a register experiment on a PG199 in which some writes succeeded and others failed; a longer write-up by @wayz2u exists, see Discord (pointer only). The PG199 is a 4096-bit A100 40GB with one HBM stack missing (32 GB), and bayley_w thinks missing stacks are restricted by a stricter mechanism. Amogh offered a test branch, and Waypoint and Wayz volunteered. Source: @wayz2u, 2026-08-29
  Source: @bayley_w, 2026-08-29
  Source: @waypoint3642, 2026-08-29
  Source: @_j3ns, 2026-08-29
• Tokukarin would prefer a DRIVE A100 board over a 10 GB 170HX (32 GB at 2.16 TB/s, 60+ TFLOPS FP16, more SMs). okrasit.: 32 GB on six stacks is suspicious, all six stacks are working, and they would unlock to 96 GB with the CMP method. Tokukarin had only heard from Chinese sellers that engineering samples unlock to 40 GB and doubts 96 GB; okrasit. countered that the 40 GB A100 has a 5 kbit bus with one stack disabled. Source: @tokukarin, 2026-09-23
  Source: @okrasit., 2026-09-23
  Source: @lia_oc, 2026-09-23

What would settle it. A PG199 showing more than 32 GB and passing a memory test.
Board. T-175 and T-250 (CMP testing) on the board.
More: 32 GB cards (subsystem 10de:1586), Other CMP cards and other NVIDIA GPUs.

Contested

Cooling and power

The 300 W VBIOS burned a card

Asked about running the 300 W VBIOS, @xanthefin said a picture of a burned 170HX had already been posted on the Discord, and, from memory, that the PCB is missing power-delivery parts. The picture has not been found in the export. The 300 W VBIOS exists for the 8 GB card only.
What would settle it. The original picture and post, with the card's VBIOS and power setup.
Board. D-090 (Dev and research) on the board.
More: Power and cabling warnings#The 300 W VBIOS has burned a 170HX.

Theory 8 GB
Source: @xanthefin, 2026-10-18
Source: @xanthefin, 2026-10-18

A 300 W VBIOS for the 10 GB card

In the order posted:

• btakk_ak47: 10 GB owners can flash a VBIOS that enables 300 W, linking TechPowerUp VGA BIOS 268495 (a CMP 170HX VBIOS). Source: @btakk_ak47, 2026-07-24
• chris.via.egnatia was buying 300 W cards that, per the seller, combine with the unlock; ju5t5omedude asked whether that "miner VBIOS" was for 8 GB cards and got no answer. Source: @chris.via.egnatia, 2026-07-24
  Source: @ju5t5omedude, 2026-07-24
• SealClubber, answering Xeax: a 300 W BIOS on the 10 GB card is not possible. Source: @xeax, 2026-08-12
  Source: @sealclubber, 2026-08-12
• MaitOPS, a 10 GB owner, believes no 300 W VBIOS exists for the 10 GB card, and Lycorascence (also 10 GB) accepted that; the stock limit defaults to 250 W. ST noted some 170HX VBIOSes draw 300 W, and DaLZy's was the 1728 MHz HBM VBIOS with the maximum power limit. Source: @lycorascence, 2026-08-25
  Source: @stcord, 2026-08-25
  Source: @dalzyyy, 2026-08-25
  Source: @_maitops, 2026-08-25
• Zagitta: the 300 W VBIOS exists only for the 8 GB card, per the Consensus-Protocol docs. Source: @zagitta, 2026-10-12
• Humvee55 runs his 128 SM 10 GB cards with the 300 W BIOS curve from the 8 GB overclock BIOS swapped in, and no extra power phases or limiter; Miséricorde claimed the card "already has a 500w bios". Source: @scallion_9883, 2026-10-16
  Source: @wakeupmbrc1, 2026-10-16

Nobody posted a measurement, and the 300 W VBIOS has burned a card (entry above).
What would settle it. A 10 GB card on a named VBIOS drawing more than 250 W sustained in nvidia-smi, with its power cabling stated.
Board. T-014 (CMP testing), D-090 (Dev and research) on the board.
More: Flashing the 300 W VBIOS, Undervolting and power limits.

Contested 10 GB

Populating the missing VRM phases allows 300 W or more, or makes the card more stable

In the order posted:

• jiacheng_liu thinks the stock VRM could handle 500 W, since the missing stages are only unpopulated, and that full TDP should be a simple shunt mod. Far checked the power stages: MP86957, rated 70 A each. Arcanum: the important missing phase is Vmem, Vcore should be fine. Luna concluded the VRM mod is pointless for core power and wanted to explore a 400 to 500 W VBIOS. Source: @jiacheng_liu, 2026-07-19
  Source: @_arcanum_, 2026-07-19
  Source: @farawaythrower, 2026-07-19
  Source: @lunaroseoc, 2026-07-19
• DaLZy likewise noted the card could be shunt-modded; nobody reported trying it. Source: @dalzyyy, 2026-08-01
• CUTE described a mod that adds the unpopulated VRM phases so cards can run at 300 W or more; he skipped it because he limits his cards to 250 W for efficiency, and said guides with parts lists are in another Discord server. Source: @cute_sc2, 2026-09-02
  Source: @humandrone8721_60165, 2026-09-02
• rickflare asked whether the MOSFETs and straps offered by a paid mod service make a difference. Civiloid: unless you reconfigure the controller IC for the power stages, adding them only wastes parts, and a service claiming the extra stages make the GPU "more stable" only has an effect if it also reprograms the controller. Arcanum: just do the caps and water-cool it. Source: @.rickflare, 2026-09-16
  Source: @_arcanum_, 2026-09-16
  Source: @civiloid, 2026-09-16
  Source: @ghostinthedatacenter_87281, 2026-09-16

What would settle it. One card with the phases populated, measured against itself before the mod at the same power limit.
Board. D-090 (Dev and research) on the board.
More: Power and cabling warnings.

Contested 8 GB and 10 GB

An unreleased "high power mode"

@matstatman: "I found a high power mode that you can enable, even on 8gb. No idea what it does so will not release." Nothing more has been posted.
What would settle it. matstatman's description of the setting and a measurement with it on and off.
Board. D-090 (Dev and research) on the board.
More: 8 GB card#Limits, Undervolting and power limits.

Theory 8 GB and 10 GB
Source: @matstatman, 2026-10-20

The 95 C slowdown can be disabled

matstatman reported on 2026-09-20: "It seems it is possible to disable the 95 degree throttling". No method, numbers or check were posted. HBM bit flips have been reported at 95 C, so disabling a thermal limit risks the card.
What would settle it. The method, a clock-event log showing the card past 95 C without slowing, and a memory-error test at that temperature. Not recommended.
Board. T-042 and T-231 (CMP testing) on the board.
More: Thermal limits.

Reported 8 GB and 10 GB
Source: @matstatman, 2026-09-20

Indium solder under the IHS, and whether delidding helps

In the order posted:

• meatsus, after a failed delid attempt, said that if it were tin or indium solder there would be no reason to delid, then cited Google Gemini saying the A100 40GB uses indium solder: an AI answer, not a teardown. Source: @meatsus, 2026-08-01
  Source: @meatsus, 2026-08-03
• quark2world: the IHS is held by perimeter glue, not by the thermal interface, and the material under it is paste. meatsus said LLMs had wrongly told him it was tin-indium solder and that the paste is about 6 years old; DaLZy found the glue rock hard; Merom Green linked a post by another user who popped the IHS. Source: @meromgreen, 2026-08-21
  Source: @quark2world, 2026-08-21
  Source: @dalzyyy, 2026-08-21
  Source: @meatsus, 2026-08-21
• SkirtChaser tried every paste he had on top of the IHS with no temperature change and concluded the bottleneck is paste under the IHS; ju5t5omedude added that several people have delidded cards by accident while removing the heatsink, which also points to no solder. Source: @skirtchaser, 2026-09-07
  Source: @ju5t5omedude, 2026-09-07
• meatsus delidded a card and measured about 10 C lower on core and memory with the stock cooler sitting on the bare die (at 150 W: about 65 C core and 75 C memory with the original paste, 55 to 60 C and 60 to 65 C with no IHS). Finishtilt14 said a delid mattered more than his own repaste. Source: @meatsus, 2026-09-15
  Source: @finishtilt14, 2026-09-16
• Miséricorde: someone already delidded and found paste, not solder, but delidding will not help because the die is huge, so thermal resistance matters less. GlobalApathy: normally IHS-to-die is indium solder. Source: @wakeupmbrc1, 2026-10-12
  Source: @globalapathy, 2026-10-12
• xanthe_: most of the thermal problem is that the paste under the IHS is old. Source: @xanthefin, 2026-10-18

What would settle it. Before-and-after core and HBM temperatures on more cards, with the same cooler and power limit, for a delid against a repaste on the IHS.
More: Thermal interface: paste, pads and phase-change, Disassembly and cleaning.

Contested 8 GB and 10 GB

The Bykski A100 block also fits the A30 and the Tesla L40

On 2026-08-21 HeadSpin quoted an unattributed claim that the Bykski N-TESLA-A100-X-V2 is the only commercial full-cover block confirmed for the 170HX, and that it also fits the A100 40 GB PCIe, the A30 24 GB and the Tesla L40. The 170HX and A100 40 GB fit is confirmed by several owners; nobody confirmed the A30 or L40 part. On 2026-08-31 Inf cited the Consensus-Protocol repo's cooling notes, which say the N-TESLA-A100-X-V2 fits and the N-TESLA-A100-80G-X-V2 does not, and RH-3D confirmed that.
What would settle it. A photo of the block mounted on an A30 or an L40.
More: Water cooling.

Theory 8 GB and 10 GB
Source: @headspin7102, 2026-08-21
Source: @inf, 2026-08-31
Source: @rh3dai, 2026-08-31

A fin splitting the duct funnel gives 30 to 50% more air pressure

kellertreppe123 (2026-09-25) runs a 3 A blower per card at about 60 C under load. For 180-degree duct adapters he said that splitting the funnel in half with a small fin improves the airstream, and claimed 30 to 50% more air pressure. Nothing was measured.
What would settle it. Static pressure, or card temperature at a fixed fan speed and power limit, with and without the fin.
More: Fans and ducts, Printable shrouds and ducts.

Theory 8 GB and 10 GB
Source: @kellertreppe123, 2026-09-25

Arctic fans wired onto the stock shroud are quieter

@danske1 has heard of people wiring Arctic fans onto the card and that "it's much quieter", but does not know what the wiring involves or how the temperatures compare. Nobody has posted a build.
What would settle it. Noise and core and HBM temperatures at a fixed power limit, before and after.
Board. T-048 (CMP testing) on the board.
More: Fans and ducts.

Theory 8 GB and 10 GB
Source: @danske1, 2026-10-26
Source: @danske1, 2026-10-26

Recipes and speed

4.5M tokens of KV cache for GLM-5.3-Flash on 4 cards

The figures differ because the setups differ: quant, KV dtype, layout and concurrency. In the order posted:

• DeeJay, on 4 cards at PCIe 2.0 x4 with his own W4A16 quant at PP4 with MTP, reported a KV pool of 4,543,829 tokens (4 concurrent requests at 1M). He then corrected that his 4,000 tok/s prefill was with BF16 KV, which gives only about a 2M pool; FP8 KV doubles the pool but drops prefill to about 2,200 tok/s. Source: @deejay9983, 2026-08-26
• Con3po, on 6 cards of an 8-card box with the NVFP4 checkpoint (W4A16 experts) at PP3 x TEP2 and MTP 3: KV 2.42M tokens; PP4 x TEP2 over 8 cards gave 4.46M to 4.82M. Source: @con3po7016, 2026-08-28
• greencandle, on 4 cards with wtdcode/GLM-5.3-Flash-AWQ-W4A16 at PP4 with MTP and no KV quantization: KV pool 1,582,946 tokens, not maximized. Source: @greencandle9753, 2026-09-09
• Miséricorde: "with 4 gpus you have w4a16 + 4.5m kv". Source: @wakeupmbrc1, 2026-10-16
• @cibernox, running the 4-card recipe daily (TP4, BF16 KV): "I only have 1.2M". Source: @cibernox, 2026-10-16

Settled. 4.5M matches DeeJay's own W4A16 build at PP4 on 4 cards; 1.2M is the TP4 recipe with BF16 KV. Quote KV capacity with the quant, KV dtype, layout and maximum sequence count, or it cannot be compared.
Board. T-069 (CMP testing) can close.
More: GLM 5.3 Flash on 4x 8 GB#KV cache and context.

Reported 8 GB

225 tok/s on Qwen 3.8 27B with one card

In the order posted:

• @feralzi: "I was able to reach 225 tks on 3.8 27b". Source: @feralzi, 2026-10-14
• @matstatman: "Its dflash and temp set to 0 i think, nothing you want to code with". Source: @matstatman, 2026-10-14

What would settle it. The same run with and without the draft model, at temperature 0 and at a normal temperature.
Board. T-158 and T-034 (CMP testing) on the board.
More: Qwen3.8 27B on 1x 8 GB (W4A16 vLLM)#Measured numbers.

Contested 8 GB

Single-card Qwen 3.8 Flash Next numbers on a custom vLLM build

In the order posted:

• @exile_1ck posted screenshots of a single 64 GB card running a Q3 quant of Qwen 3.8 Flash Next on a "custom vllm build aimed at the 170hx". Source: @exile_1ck, 2026-10-15
  Source: @exile_1ck, 2026-10-15
• @xtj7 gets about 90 tok/s decode and 1600 tok/s prefill with Strata on one card at IQ2 without the x16 mod, called eXile's prefill "awfully high" and wondered whether it includes cache hits. @okrasit. asked whether the prefill was reported correctly, @shudza asked for the context size. No answers were posted. Source: @xtj7, 2026-10-15
  Source: @okrasit., 2026-10-15
  Source: @xtj7, 2026-10-15
  Source: @shudza, 2026-10-15

What would settle it. The numbers with context size, system RAM and cache hits stated, and the build published.
Board. T-077 (CMP testing) on the board.

Contested 8 GB

v0.5 decodes slower than v0.4

In the order posted:

• @reala7096: ECC has no effect on bandwidth. @matstatman: random one-byte writes would suffer. Source: @reala7096, 2026-10-05
  Source: @matstatman, 2026-10-05
• @mowgli8628: after reinstalling with v0.5 he "can't get past 35 t/s with qwen 3.8 27b", against 60-80 on v0.4, and asks whether ECC is the cause. Source: @mowgli8628, 2026-10-18

What would settle it. The same model and settings on v0.4 and v0.5, with ECC on and off, on one card.
Board. T-003 (CMP testing) on the board.
More: Driver and unlock failures#Performance changed after an update.

Contested 8 GB

Broadwell CPUs halve the Qwen Flash recipe numbers

In the order posted:

• @segmond reached nowhere near Morrowmake's GLM numbers and put it down to missing x16 links and TP. Source: @segmond, 2026-10-20
• @nathang6866 (ElectroHobo) got about half of Morrowmake's numbers on the Qwen Flash recipe: "It seems the Broadwell processors cause issues with certain Qwen Flash recipes." Source: @nathang6866, 2026-10-20

What would settle it. The same recipe on a Broadwell host and a newer host with the same cards, links and P2P setting.
Board. T-065 (CMP testing) on the board.
More: Platforms#Host CPU and RAM.

Contested 8 GB

Buying

One card earns up to 7,000 USD a year

In the order posted:

• @cute_sc2: "a single cmp 170hx can now earn up to 7k a year in profit alone, you only need to know how", through long-term rental contracts with cheap energy, Pearl mining or serving AI. Source: @cute_sc2, 2026-10-23
  Source: @cute_sc2, 2026-10-23
• @kellertreppe123: that is a service business, not the card's hardware value, and businesses would rather trust a cloud provider with a track record. Source: @kellertreppe123, 2026-10-23
  Source: @kellertreppe123, 2026-10-23

What would settle it. An owner's income and costs per card over several months.
More: Price history and market.

Contested 8 GB and 10 GB

Half of all cards are bad

In the order posted:

• @_____josh_____: "the cards have a 50% failure rate". Source: @_____josh_____, 2026-10-08
• @lia_oc: "It's nowhere near that high for sure". Source: @lia_oc, 2026-10-08
• @c133 tested five cards, all good; @bigfoot1144 bought 12, all unlock. Source: @c133, 2026-10-08
  Source: @bigfoot1144, 2026-10-08
• @_____josh_____: the good stock sold earlier and sellers now mix lemons into each shipment. Source: @_____josh_____, 2026-10-08

What would settle it. A structured survey: cards bought, source, unlock result, HBM error count, tensor-core check.
Board. T-115 (CMP testing) on the board.
More: Buying#How often cards are bad, Card failure rates and longevity.

Contested 8 GB and 10 GB

The vouches for Alifond are bot accounts

In the order posted:

• @kellertreppe123: Alifond (Huateng) is vouched for by several members here. Source: @kellertreppe123, 2026-10-04
• @_____josh_____: "The multiple guys are their own bot accounts. Don't fall for it." @djpandaman: "this is true tho". Source: @_____josh_____, 2026-10-04
  Source: @djpandaman, 2026-10-04

What would settle it. Order reports from long-standing members, with dates and what arrived.
More: Buying#Alifond.

Contested 8 GB and 10 GB

Listings promising PCIe 4.0, NVLink, ECC and 80 GB

In the order posted:

• An eBay listing showed a 170HX with PCIe 4.0. @con3po7016: "false alarm"; @spectakkular: "PCIe 4.0 is a listing mistake from before the cards were cracked. It's just eBay autofilling listing information". Source: @con3po7016, 2026-10-11
  Source: @spectakkular, 2026-10-11
• @tugoubxs found a seller whose description promised ECC, PCIe 4.0, NVLink and 80 GB. Source: @tugoubxs, 2026-10-14
  Source: @tugoubxs, 2026-10-14
• A seller's screenshot read as PCIe Gen4 x16 by @galygious2; @xeax: the field shows x1 Gen1 at idle. Source: @galygious2, 2026-10-18
  Source: @xeax, 2026-10-18

What would settle it. Nothing in a listing counts; check the card on arrival.
More: Buying#Scam patterns, Acceptance checklist.

Contested 8 GB and 10 GB

Die transplants and cards with no chips

@civiloid recalled reports of used-market cards "with no chips", and of a CMP that turned out to carry a die transplanted from the 32 GB A100 variant. No photo or link to the original reports was posted.
What would settle it. The original reports or photos.
More: Which card do I have, Buying.

Theory 8 GB and 10 GB
Source: @civiloid, 2026-10-13
Source: @civiloid, 2026-10-13

A new finding is behind the late-October price jump

In the order posted:

• Miséricorde: "seems like someone found something"; a card he bought at 2,380 USD in the CIS market cost him 3,600+ on arrival. Source: @wakeupmbrc1, 2026-10-22
  Source: @wakeupmbrc1, 2026-10-22
• @ob88594: "No, what I know is that Russians are aggressively purchasing the 8G version in China"; bulk lots of 10 cost 20% more than the regular price. Source: @ob88594, 2026-10-22
  Source: @ob88594, 2026-10-22

What would settle it. A published unlock that explains the jump; otherwise it stays a market question.
More: Price history and market.

Contested 8 GB

At least 130,000 8 GB cards exist

@spectakkular: "there seems to be at least 130,000 of the 8G around the world". No source for the figure was given.
What would settle it. A shipment or production figure with a source.

Theory 8 GB
Source: @spectakkular, 2026-10-05

Seller screenshots prove nothing

In response to the 124-SM seller claims (2026-08-12), Studebaker posted mock images (SealClubber identified one as a "cmp420" joke) and an LLM-generated script to show how easily such status screenshots are faked, and said the hype appeared when sales slowed. Only live, independently run checks (register reads, SM count, GEMM throughput) mean anything.
More: Verifying an unlock, Acceptance checklist.

Warning 8 GB and 10 GB
Source: @_studebaker_, 2026-08-12
Source: @sealclubber, 2026-08-12

The "42069 special" joke screenshot was reused as proof and in advertising

On 2026-08-17 an Alibaba seller sent Arcanum a meme nvidia-smi image (420 and 69 numbers, "Nice/sec" benchmarks) that Studebaker had made in this Discord, quoted 1,600 USD, and said they "found it on a forum" when asked for details. kellertreppe123 showed a seller raising the price by US$100 in 5 hours, and philow got a quote of 1,700 per card the same day; machinezero objected to sellers raising prices off a troll screenshot. On 2026-08-23 Studebaker reported that a seller in China had used the image (giaki3003's version claimed a "128GB unlock" and showed PCIe 4.0 x16) in advertising, presenting it as what the card is. Any listing showing 128 GB, PCIe 4.0 x16 or "42069" values for a CMP 170HX is fake.
More: Buying.

Warning 8 GB and 10 GB
Source: @_arcanum_, 2026-08-17
Source: @irythros, 2026-08-17
Source: @kellertreppe123, 2026-08-17
Source: @.philow, 2026-08-17
Source: @machinezer0, 2026-08-17
Source: @_studebaker_, 2026-08-17
Source: @_studebaker_, 2026-08-23
Source: @giaki3003, 2026-08-23
Source: @confident_lamb_62706, 2026-08-23
Source: @irythros, 2026-08-23

"PCIe 4.0 / full CUDA" cards on Xianyu: an A800 in disguise

In the order posted:

• 丁薇 posted screenshots from Xianyu of a seller claiming a 170HX with PCIe 4.0 and all CUDA cores unlocked. Source: @dingwei0049, 2026-08-12
• zapirate: the SM and CUDA core counts match older engineering-sample cards. Source: @zapirate, 2026-08-12
• turbine, in the thread with Rubbra_F: the bench showed about 10 GB/s, while Gen4 x16 should give about 32 GB/s (receive was over 16 GB/s), maybe through a PLX switch, and the OpenCL bench used does not test tensor-core throughput. Source: @rubbra_f, 2026-08-12
  Source: @turbine, 2026-08-12
• intervitens: the seller's video shows PCI IDs 0x10de 0x20bd, which is the 40 GB A800, so either an A800 renamed in the driver or an A800 BIOS put on a 170HX. Source: @intervitens, 2026-08-12
• Humvee55, after Wayz: ask for a full register dump, otherwise drop it. Source: @wayz2u, 2026-08-13
  Source: @scallion_9883, 2026-08-13

What would settle it. A full register dump from the card.
More: Verifying an unlock.

Warning 8 GB and 10 GB

A Chinese paid service for ECC and PCIe 4.0 unlocks (August 2026)

Apoloniato shared screenshots on 2026-08-10 of a Chinese poster offering a paid service to unlock ECC and PCIe 4.0 on the 170HX, judging the account credible from its history. The same poster says 10 GB cards cannot unlock to 80 GB (DaLZy). Zarknoth said his own seller advertised PCIe 4.0 but he has not seen it, and questioned who would pay the asking price; Apoloniato thought it was still cheaper than an A100 cluster. ECC later arrived for everyone in cmpunlocker v0.5 (ECC); the PCIe 4.0 part is unverified.

Reported 8 GB and 10 GB
Source: @apoloniato_07709, 2026-08-10
Source: @zarknoth, 2026-08-10
Source: @dalzyyy, 2026-08-10

A Goofish listing: a 10 GB card with a PCIe 4.0 unlock

In the order posted (2026-08-31):

• Miséricorde found a Goofish listing whose seller claims they can now unlock PCIe to 4.0 and says the card is a 10 GB CMP; he suspects it is a lie and the card is just an A100 40G (the screenshot shows 7936 cores). Source: @wakeupmbrc1, 2026-08-31
• giaki3003: "100% real", with a meme gif, so his tone is unclear. Source: @giaki3003, 2026-08-31
• Arcanum: much is wrong with the screenshot, especially the bandwidth numbers. Source: @_arcanum_, 2026-08-31

What would settle it. lspci -vv from the card on arrival.

Contested 10 GB

An eBay seller's "secret Chinese unlock is coming" pitch

peter643 reported on 2026-09-04 that at least one eBay seller, after you buy a card, sends the Chinese SM and Gen4 unlock images and says a secret Chinese unlock is coming, then stops replying when asked for details. Treat it as hype, not as a feature of the card.

Reported 8 GB and 10 GB
Source: @peter643, 2026-09-04

8 GB listings that promise Gen3 (September 2026)

On 2026-09-26 Josh asked that the Gen3 failure be announced in the main announcements channel so that "alifond bots" (apparently accounts pushing the supplier Alifond) stop scamming uninformed buyers with Gen3 claims. As of 2026-09-26 there was no released Gen3 unlock for 8 GB cards, so an 8 GB listing that promises Gen3 is false.
More: 8 GB card#Limits, Buying.

Warning 8 GB
Source: @_____josh_____, 2026-09-26

Moving a 170HX core onto an A100 PCIe board

OB88594 said on 2026-09-29 that moving a 170HX core onto an A100 PCIe board ("170HX-A100PCIE") costs an extra US$400; some people are transplanting cores to test whether NVLink works, but it is still at the engineering stage and nobody sells it yet. He expects unlocking a 10 GB card to more than 48 GB to be sold only as a complete service. Waypoint: hardly worth the labour and equipment. OB88594 compared its prospects to the early PCIe 3.0 unlock talk: far away.

Reported 8 GB and 10 GB
Source: @waypoint3642, 2026-09-29
Source: @ob88594, 2026-09-29
Source: @zingy90210, 2026-09-29

regsky_ (Greg) and gpulab.net: a paid mod service and a PCIe 3.0 claim

The X account regsky_ (called Greg in the server, known for a 4090 upgrade business) began posting about the 170HX in September 2026; ST noted that even broken cards' cores are worth money. meatsus: Greg claimed a PCIe 3.0 unlock but showed only lspci output and said he used cmpunlocker v0.3; ST pointed out that the unlock had been published in this server for some time. Arcanum linked a gpulab.net product page selling a 170HX service, apparently about US$500 (matstatman: "500 usd for caps?"), that also mentions MOSFETs and straps. The seller also said moving the package onto an A100 board would need an A100 PCB and cooler at a higher cost. Members treated the PCIe 3.0 claim as unproven.
More: the VRM-phase entry under Cooling and power, PCIe x16 capacitor mod.

Contested 8 GB and 10 GB
Source: @stcord, 2026-09-16
Source: @meatsus, 2026-09-16
Source: @zefrench6961, 2026-09-16
Source: @_arcanum_, 2026-09-16
Source: @matstatman, 2026-09-16
Source: @djpandaman, 2026-09-16

aifounder: 8 GB cards at 2,000 USD or more with the PCIe speed mod done

aidoc reported on 2026-07-22 that a seller from "aifounder" claimed to sell the 8 GB card at 2,000 USD or more with the PCIe speed mod and soldering applied. giaki3003: the miners bought these cards for about 4k five years ago, never made the money back, and are hoping for A100 pricing.

Reported 8 GB
Source: @giaki3003, 2026-07-22
Source: @sekkit_75019, 2026-07-22

A VBIOS port would bring fake "A100 64GB" cards

bayley_w (2026-08-10) warned that porting the unlocks into a VBIOS could bring fake "A100 64GB" cards out of China, and that making the cards plug and play favours resellers and could make buyers pay 2 to 3 times more. 64 replied that plug and play is good. Waka324 countered that it would filter out eBay scammers, since today a buyer cannot claim bad memory on a card advertised only as a mining card running mining loads.

Warning 8 GB and 10 GB
Source: @bayley_w, 2026-08-10
Source: @64pigs, 2026-08-10
Source: @cmdrwaka324, 2026-08-10

NVIDIA buys back GPUs at 20 to 25% of the original price

irythros (2026-09-03) relayed unverified Discord talk that NVIDIA runs a GPU buy-back program, repurchasing at 20 to 25% of the original price, but found nothing formal. CUTE agreed, "thus keeping market tight". Which GPUs this concerns is not clear.
What would settle it. A published NVIDIA programme or a seller's receipt.
More: Price history and market.

Theory 8 GB and 10 GB
Source: @irythros, 2026-09-03
Source: @cute_sc2, 2026-09-03

See also

Buying, Price history and market, Other CMP cards and other NVIDIA GPUs, 8 GB card, 10 GB card, Which card do I have, Verifying an unlock, VBIOS and firmware, Unlock history: who found what, and when, The 10 GB card: provenance and the 80 GB question, Status taxonomy, Board how-to.

Imported from the cmpunlocker Discord with credit; see Credits.