Rumours and claims to verify
More actions
Claims made on the cmpunlocker Discord that nobody has reproduced. None of them is a result. 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.
- 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.
- 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:
- 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
- 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:
- 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 and T-027 (CMP testing), D-055 (Dev and research) on the board.
More: Memory unlock on the 10 GB card#80 / 96 GB: not real on these dies.
Contested 10 GB
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
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.
Contested 8 GB
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
- Ivan's 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.
Contested 8 GB
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 - Ivan's 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
PCIe 5.0 on the 170HX
In the order posted:
- 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).
Contested 8 GB and 10 GB
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
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 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
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
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
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
In the order posted:
- Miséricorde: "with 4 gpus you have w4a16 + 4.5m kv". Source: @wakeupmbrc1, 2026-10-16
- @cibernox, running the 4-card recipe daily: "I only have 1.2M". Source: @cibernox, 2026-10-16
What would settle it. The vLLM startup line together with KV dtype, layout and maximum sequence count.
Board. T-069 (CMP testing) on the board.
More: GLM 5.3 Flash on 4x 8 GB#KV cache and context.
Contested 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.
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