The 10 GB card: provenance and the 80 GB question

The 10 GB card (device 10de:2082, Samsung HBM on five active stacks) unlocks to 40 GB. From July to early October 2026 owners and developers tried to take it to 80 GB, the figure its bus width and stack count seem to promise. Every attempt hit the same wall: 80 GB shows up in nvidia-smi, and kernels can even write and read much of it, but loading, copying or freeing data above about 40 GB crashes the program, the driver or the GSP firmware. This page keeps that history, and where the cards came from, so that 10 GB card and Memory unlock on the 10 GB card can stay short. For what to do with a 10 GB card today, start on those pages.

Where it stands: 40 GB is the stable unlock (Proven). 48 GB is claimed for some cards whose sixth stack is intact (Contested). 80 GB has not been achieved: it can be made visible, not used. Every claim of a working 80 GB below is marked Contested.

Where the cards came from

Production weeks: 8 GB cards from 2021, 10 GB cards from 2022

CMP serials follow one pattern: the prefix 132, a two-digit week, a two-digit year, then a six-digit sequence. On 2026-08-18 Impactic collected serials posted in this server and in a clean-room server: 35 confirmed 8 GB cards came from weeks 22, 23, 24, 26, 28 and 29 of 2021, and 3 confirmed 10 GB cards from week 04 of 2022. RailGun (4 cards), Toffles (2) and Lycorascence (1) also posted week-04-2022 serials for 10 GB cards, and cuddylac997's four 10 GB cards show weeks 02, 03 and 04 of 2022. The 132 prefix also appears on RTX 3090, A6000, A5000, T4 and L4 boards, and Studebaker thinks the sequence counts all GA100 chips, so it is not a CMP production count. Read your own serials with nvidia-smi --query-gpu=index,serial --format=csv,noheader. RailGun left off his final digit for privacy, and this page lists no full serials.

Reported 8 GB and 10 GB
Source: @impactic, 2026-08-18
Source: @_studebaker_, 2026-08-18
Source: @lycorascence, 2026-08-18
Source: @railgun9987, 2026-08-18
Source: @toffles3179, 2026-08-18
Source: @cuddylac997, 2026-08-18
Source: @gk_hun, 2026-08-18

Where each variant came from: A100 80GB dies or A100 40GB stock (contested)

Members tie the 8 GB card to the A100 80GB (SK Hynix memory) and the 10 GB card to the A100 40GB (Samsung memory). They disagree on the production order and on whether the 10 GB card's hidden half is failed or only switched off. In the order posted:

• 2026-07-25: @dalzyyy held that the 10 GB card's Samsung memory is binned 80 GB-class HBM2e sold as 40 GB HBM2; @rrickk420 guessed that SK Hynix simply has better QA than Samsung (an untested opinion). Source: @giaki3003, 2026-07-25
  Source: @dalzyyy, 2026-07-25
  Source: @rrickk420, 2026-07-25
• 2026-08-06 and 08-07: @skirtchaser noted that the A100 40GB has the full 5120-bit bus with 40 GB accessible, like the unlocked 10 GB card, so the 10 GB card likely comes from A100 40GB stock with five 8 GB stacks. @bayley_w added that the A100 40GB uses HBM2 and the 80GB HBM2e with double stack height and higher clocks, so the 10 GB card's lower memory clock fits HBM2; @giaki3003 first thought HBM2e, then agreed that more than 40 GB is impossible. For the 8 GB card SkirtChaser proposed an A100 80GB die with one stack or memory controller failed, leaving 4x 16 GB = 64 GB. @robbbbbb7973's 8 GB card shows 1728 MHz memory in nvidia-smi (narrower bus, higher clock), @_studebaker_ linked the 170tune HBM matrix and recapped: 10 GB = A100 40GB old stock, 8 GB = A100 80GB made in production. @sealclubber said 10 GB supply is lower. Source: @bayley_w, 2026-08-06
  Source: @skirtchaser, 2026-08-06
  Source: @giaki3003, 2026-08-06
  Source: @robbbbbb7973, 2026-08-06
  Source: @_studebaker_, 2026-08-06
  Source: @sealclubber, 2026-08-06
  Source: @skirtchaser, 2026-08-07
  Source: @_studebaker_, 2026-08-07
  Source: @giaki3003, 2026-08-07
• 2026-08-09: bayley_w called the 8 GB card an A100 80GB with two bond failures, and DaLZy pointed to a since-deleted post that labelled 2 FBPs disabled and 2 defective on an 8 GB card, which he took as confirmation. The same day's thread on six-stack cards is under #48 GB and the sixth stack. Source: @dalzyyy, 2026-08-09
  Source: @bayley_w, 2026-08-09
  Source: @bayley_w, 2026-08-09
  Source: @aaasss1456, 2026-08-09
  Source: @dalzyyy, 2026-08-09
  Source: @k2v0008, 2026-08-09
  Source: @_____josh_____, 2026-08-09
  Source: @hastwell, 2026-08-09
  Source: @skirtchaser, 2026-08-09
• 2026-08-11: @petrikrohn_26497: the 8 GB card is an A100 80GB with one HBM2 stack disabled by fuses, and the 10 GB card an A100 40GB whose Samsung stacks are reportedly rejected 16 GB stacks sold by Samsung as 8 GB. @cmdrwaka324 had asked whether the cards were unsold A100 dies, and @phoenixone7 added that fusing off HBM is physically difficult, which is why the memory was still there to unlock. None of this came with sources. Source: @petrikrohn_26497, 2026-08-11
  Source: @cmdrwaka324, 2026-08-11
  Source: @phoenixone7, 2026-08-11
• 2026-08-12: giaki3003 said production dates show the 10 GB card was made last, from the first binned A100 40GB dies, so it cannot carry 80 GB; two days earlier he had said the 8 GB cards were the newest and that the 10 GB cards were built from older parts once stock ran out. @meatsus countered that the 8 GB card could not exist without being cut down from the 80 GB design. Source: @giaki3003, 2026-08-12
  Source: @meatsus, 2026-08-12
• 2026-08-16: bayley_w described the 10 GB cards as late-production A100 40G that used Samsung 16 GB HBM2e stacks downbinned to 8 GB HBM2 stacks for yield, leftover A100 40G resold as CMP 170HX after the A100 80G shipped, and concluded that the top 40 GB is guaranteed broken. Source: @bayley_w, 2026-08-16
  Source: @lunaroseoc, 2026-08-16
  Source: @arabus_74064, 2026-08-16
• 2026-09-16: @cute_sc2's fuse readouts point the other way for the 8 GB card: all six dies are present on most cards, deactivated rather than failed. Source: @cute_sc2, 2026-09-16
  Source: @quark2world, 2026-09-16
  Source: @segmond, 2026-09-16

No die photo or cross-section was posted. The Chinese kkk.rs write-up reports physically 16 Gb dies on the 10 GB card (#Reading guides from China).

Contested 8 GB and 10 GB

A100 and CMP timeline, and the CoWoS argument

bayley_w (2026-08-09) put the A100 40G at May 2020, the A100 80G in practice at Q1 2021, Samsung HBM2e at the first half of 2020 and the CMP 170HX at September 2021. He argued that NVIDIA would not spend scarce CoWoS packaging capacity on a $4299 card instead of a $10K card, because GA100 dies sat unpackaged at the time, and found it plausible that the 10 GB card is leftover A100 40GB stock from after the 80GB launch. He also said GA100 engineering samples were "a zoo" with 32, 64 and 96 GB configurations. DaLZy replied that in 2021 and 2022 NVIDIA could not sell more A100s at $10K, and that binning is often an economic decision rather than a matter of defects. Nobody answered bayley_w's question of when the 10 GB SKU first appeared on the mining market; the serial survey above puts 10 GB production in weeks 02 to 04 of 2022.

Theory 8 GB and 10 GB
Source: @bayley_w, 2026-08-09
Source: @dalzyyy, 2026-08-09
Source: @phoenixone7, 2026-08-09

Why the CMP 170HX exists: failed A100 dies, unsold A100 silicon or market segmentation (contested)

In the order posted:

• 2026-07-20: @lunaroseoc said the cards came from a bad batch of A100 dies with dead memory and dead cores that could not be sold as 80 or 96 GB parts, so half the chip was cut off and sold to miners. She also said an A100 64GB never existed (bayley_w later listed a 64 GB engineering sample). Prices given then: the CMP 170HX at launch about $4300 (@a.civardagezen), the A100 40GB about $10K and the 80GB about $20K (Luna, from memory). Source: @lunaroseoc, 2026-07-20
  Source: @a.civardagezen, 2026-07-20
  Source: @_arcanum_, 2026-07-20
  Source: @jonathan.quai, 2026-07-20
• 2026-09-09: @_____josh_____ said the dies are binned chips that failed QC. @cmdrwaka324 countered that yield on the 8 GB HBM dies was fine and that NVIDIA used the CMP to offload unsold A100s without cannibalising A100 sales; @spectakkular added that they were sold to China after A100 exports were banned; @quark2world noted that NVIDIA made millions of these chips. Source: @_____josh_____, 2026-09-09
  Source: @cmdrwaka324, 2026-09-09
  Source: @spectakkular, 2026-09-09
  Source: @quark2world, 2026-09-09
• 2026-09-16: Josh asked what business sense it made to sell a working A100 for $1000 instead of $10000. @dalzyyy answered price discrimination (the CMP originally sold for about $4000), quark2world said mining cards sold for 5 to 6k while a new A100 was closer to 30k, and @cute_sc2 pointed to A100 oversupply that could not be sold at the set price without cannibalising the datacenter market. Source: @_____josh_____, 2026-09-16
  Source: @dalzyyy, 2026-09-16
  Source: @quark2world, 2026-09-16
  Source: @cute_sc2, 2026-09-16
• 2026-09-18: @xtj7: at the end of the mining boom NVIDIA turned 16 GB V100s into the CMP 100-210 and sold A100 silicon as the CMP 170HX, still profitable with that much memory capped off; he calls it an anomaly unlikely to repeat. @ronnyfish added that memory chips cost low single-digit dollars then. Source: @xtj7, 2026-09-18
  Source: @ronnyfish, 2026-09-18
• 2026-09-20: Josh claimed the cards carry known-defective, lower-binned memory from batches that failed QC. @wakeupmbrc1 disputed it: they are the same GA100s on the same boards as A100s, production A100s also ship with disabled GPCs, disabled GPCs say nothing about lifespan, and the CMP SKU was "artificial segmentation". Source: @_____josh_____, 2026-09-20
  Source: @wakeupmbrc1, 2026-09-20

Neither side offered evidence beyond assertion.

Contested 8 GB and 10 GB

Open question: why does the 10 GB card keep five stacks?

mrpandagod asked on 2026-09-27 why NVIDIA left five stacks enabled on the 10 GB card, unchanged from the A100 40GB, but fused off one more stack on the 8 GB card than on the A100 80GB. The only reply was a joke.

Theory 8 GB and 10 GB
Source: @djpandaman, 2026-09-27

Theory: the 170HX boards are PG199 parts NVIDIA could not sell

DaLZy (2026-09-16) noted that a PCIe version of the PG199 apparently exists and suggested "all those pg170hx are probably just pg199 they couldn't sell". Later in that chat the PG199 is described as a 32 GB SXM2 part selling for about $1,500, with third-party SXM2 adapters available. An unverified idea; members identify the 32 GB-only boards sold as 170HX as PG199 parts (32 GB cards (subsystem 10de:1586)).

Theory
Source: @dalzyyy, 2026-09-16

HBM date codes predate the boards by three months or more

matstatman found the HBM about three months or more older than the board (2026-09-20), which he framed as normal procurement, since memory is bought ahead of assembly; giaki3003 said it fits the theory that these stacks were made for 80 GB A100s and ended up on mining cards. On 2026-09-16 Toffles read a die identity frame whose date points to early HBM2e or late HBM2, while 10 GB serials point to week 4 of 2022. His three hypotheses: a mixed HBM2 and HBM2e batch given one VBIOS that suits the slower part; Samsung tried 16 GB stacks and had to bin them all down to 8 GB; or the original 80 GB claim was wrong. He asked owners to share the manufacture dates from their own die identity frames, and MaitOPS and 68x pointed him to the kkk.rs write-up.

Theory 10 GB
Source: @toffles3179, 2026-09-16
Source: @_maitops, 2026-09-16
Source: @68x0, 2026-09-16
Source: @matstatman, 2026-09-20
Source: @giaki3003, 2026-09-20

Memory vendor and bus width: Samsung on 5120 bits, SK Hynix on 4096 bits

On 2026-07-19 Mr Pec stated that all 10 GB cards use Samsung HBM and RazK confirmed that all 8 GB cards are SK Hynix and all 10 GB cards Samsung; Luna and supasnipa360 repeated it the next day. Thanos counted 4 working stacks on the 8 GB card and 5 on the 10 GB card, and coolgod derived 4096-bit and 5120-bit buses and, at 16 GB per stack, 64 GB and 80 GB. Arcanum: the 8 GB card's HBM runs at about 1458 MHz and the 10 GB card's at about 1215 MHz at HBM2 speeds, so the extra stack gives bandwidth parity; TechPowerUp lists 1.49 TB/s (8 GB) and 1.56 TB/s (10 GB). Archimedes compared the A100 80GB (1.94 TB/s at 1512 MHz) with the 170HX (1.59 TB/s at 1215 MHz). Mr Pec added that the 8 GB card was always the better mining card. Kidako (2026-07-30) runs four "8GB Hynix" cards and three "10GB Samsung" cards, and Rickcal (2026-09-16) said the two behave much the same apart from the memory vendor. Whether the packages carry more stacks than are enabled is a separate question (#48 GB and the sixth stack).

Proven 8 GB and 10 GB
Source: @mrpec, 2026-07-19
Source: @port00, 2026-07-19
Source: @coolgod_72592, 2026-07-19
Source: @thanos8912, 2026-07-19
Source: @coolgod_72592, 2026-07-19
Source: @meatsus, 2026-07-19
Source: @meme_meme, 2026-07-19
Source: @lunaroseoc, 2026-07-20
Source: @supasnipa360, 2026-07-20
Source: @_arcanum_, 2026-07-20
Source: @archi.medes, 2026-07-20
Source: @giaki3003, 2026-07-25
Source: @dalzyyy, 2026-07-25
Source: @rrickk420, 2026-07-25
Source: @kidazuk, 2026-07-30
Source: @_arcanum_, 2026-07-30
Source: @moliminous, 2026-07-30
Source: @bayley_w, 2026-08-16
Source: @lunaroseoc, 2026-08-16
Source: @arabus_74064, 2026-08-16
Source: @napsterae, 2026-09-16
Source: @rrickk420, 2026-09-16

10 GB supply: bulk buying in China and large holders (unverified)

On 2026-09-10 Takka and ST linked X posts by Defi_Django, apparently a large CMP holder (one post was old, per matstatman; DaLZy said "They still have 1500 apparently"). Humvee55 said he bought 20 10 GB cards from that seller before the price spike, and cuddylac997 claimed to hold thousands of 10 GB cards and to be holding off selling after learning they were not 8 GB cards. On 2026-09-19 OB88594 said the 10 GB version was being bought in bulk on the Chinese market through specific channels (Xianyu), with no further details. All of this is unverified, and no prices were given. Prices over time: Price history, July to October 2026.

Reported 10 GB
Source: @spectakkular, 2026-09-10
Source: @dalzyyy, 2026-09-10
Source: @stcord, 2026-09-10
Source: @scallion_9883, 2026-09-10
Source: @matstatman, 2026-09-10
Source: @cuddylac997, 2026-09-10
Source: @ob88594, 2026-09-19

How many cards were made: 12,000 to 225,000 (contested)

Estimates of the production run, in the order posted:

• 2026-07-29: @dalzyyy estimated over 30,000 units (no source given); @machinezer0 said there are not enough 170HX cards to move other GPU prices. Source: @dalzyyy, 2026-07-29
  Source: @machinezer0, 2026-07-29
• 2026-08-03: @mrwojaketh said 330k were made, citing the internet, then corrected himself ("Nvm, cmp lineup"), so that figure probably covers the whole CMP line. Source: @mrwojaketh, 2026-08-03
  Source: @cute_sc2, 2026-08-03
• 2026-08-06: DaLZy put it at 30k to 100k+; giaki3003 cited a paper estimating 12k to 36k, which DaLZy called guesswork based on a share of revenue and sensitive to the bulk price ($3k against $4k to $5k); @quark2world thought the CMP line existed because NVIDIA had not yet sold enough of the datacenter part; @xeax expected CMP prices to rise with GPU prices. Source: @dalzyyy, 2026-08-06
  Source: @giaki3003, 2026-08-06
  Source: @quark2world, 2026-08-06
  Source: @xeax, 2026-08-06
• 2026-08-10: DaLZy said that if one customer could buy about 2800 the total may be near 100k; @scallion_9883 said one person reportedly holds 8k alone. Source: @dalzyyy, 2026-08-10
  Source: @scallion_9883, 2026-08-10
  Source: @cuddylac997, 2026-08-10
• 2026-08-20: giaki3003 estimated 12k to 36k; @bentbean_50370, assuming rejected A100 dies, guessed at most 50,000; @_studebaker_ doubted that serial suffixes reveal production numbers and called every figure a guess; @sealclubber said there are a lot of CMPs and anyone claiming otherwise is selling; quark2world's eBay sales data suggested only about 3k cards a year turn over there. Source: @giaki3003, 2026-08-20
  Source: @bentbean_50370, 2026-08-20
  Source: @_studebaker_, 2026-08-20
  Source: @sealclubber, 2026-08-20
  Source: @quark2world, 2026-08-20
  Source: @adusts, 2026-08-20
• 2026-08-30: @segmond guessed 225,000. Studebaker disputed it, since serial numbers are shared across all GA100 products and his highest suffix is 129,941; @danske1's lowest and highest were 000,925 and 128,617, @.philow guessed about 20k (with 225k cards prices would be below $1k), Studebaker knows people who together hold about 12k cards, and segmond noted that big buyers bought hundreds at a time, so supply is concentrated. Source: @segmond, 2026-08-30
  Source: @_studebaker_, 2026-08-30
  Source: @danske1, 2026-08-30
  Source: @.philow, 2026-08-30
  Source: @reala7096, 2026-08-30
• 2026-10-05: @spectakkular estimates at least 130,000 8 GB cards exist worldwide. @_____josh_____ argued that the unlocked memory is lower-binned, failure-prone memory NVIDIA fused off; @wakeupmbrc1 disputed it (every card runs at ndiv=68 as a minimum), and @jjb01957's card runs fine. Neither side gave failure statistics. Source: @spectakkular, 2026-10-05
  Source: @_____josh_____, 2026-10-05
  Source: @wakeupmbrc1, 2026-10-05
  Source: @jjb01957, 2026-10-05

Contested 8 GB and 10 GB

40 GB from the first day

Earliest reports: 40 GB guaranteed, 80 GB unstable (June and July 2026)

On 2026-06-22, in the server's earliest history, atsasin passed on second-hand information: the 10 GB card works fine at 40 GB, 80 GB had been reached on it but was very unstable and had to run at half speed, and the 8 GB card ran stable at 64 GB. He put the instability of the upper layers down to the Samsung HBM (the 8 GB card uses SK Hynix), said he had not seen it himself and asked for replication; MrLoki89 had suggested that the fifth stack might be faulty on some cards. On 2026-07-14, replying to the unlock release, Petri Krohn wrote: "Only 40GB is guaranteed to work on the 10GB card with Samsung stacks." Only these short claims are summarised here; the register values in the 2026-07-14 message are not reproduced.

Reported 10 GB
Source: @atsasin2126, 2026-06-22
Source: @serplay228, 2026-06-22
Source: @petrikrohn_26497, 2026-07-14

July 2026 baseline: 8 GB to 64 GB clean, 10 GB to 40 GB stable, 80 GB not

On 2026-07-19 CUTE asked whether the 64 GB cards show any HBM errors and Amogh answered "Nope, none at all." Far ran cuda_memtest on a 10 GB card with the 80 GB profile: it ran when allocating 39 GB and failed at 40 GB or more, while "40GB has worked perfectly stable" for him. coolgod summed up that instability had been seen only on the 80 GB unlock, not at 40 GB and not on 8 GB cards at 64 GB. Eksitus (2026-07-20) said he and another tester agreed that using under 40 GB is fine on the 80 GB profile, but any attempt to use the rest crashes the running process. Luna wrote on 2026-07-20 that 80 and 96 GB "cause errors and/or no POST at the moment" and "Right now it's safe to say 8 -> 64 and 10 -> 40."

Proven 8 GB and 10 GB
Source: @cute_sc2, 2026-07-19
Source: @_j3ns, 2026-07-19
Source: @farawaythrower, 2026-07-19
Source: @coolgod_72592, 2026-07-19
Source: @farawaythrower, 2026-07-20
Source: @lunaroseoc, 2026-07-20
Source: @eksitus, 2026-07-20

40 GB worked from the first unlock; 80 GB showed but was not addressable

On 2026-07-19 Far reported his 10 GB cards (10de:2082) working properly at 40 GB, though without PCIe Gen2, and GMZombie007 had two 10 GB cards at 40 GB under HiveOS. Unlocking to 80 GB made the cards report 80 GB, but the memory was not fully addressable and benchmarks did not pass consistently. By 2026-07-21 the agreed state (Arcanum, Waypoint, Luna) was 8 GB to 64 GB stable, 10 GB to 40 GB stable, and 80 GB on the 10 GB card possible but unstable; supasnipa360 gave 10 GB to 80 GB as a target, not a result. meatsus and giaki3003 noted that the VRAM refresh-rate change from Jon Pry's paper can be applied, but the result is not stable enough.

Proven 10 GB
Source: @farawaythrower, 2026-07-19
Source: @gmzombie, 2026-07-19
Source: @lunaroseoc, 2026-07-19
Source: @_arcanum_, 2026-07-21
Source: @waypoint3642, 2026-07-21
Source: @supasnipa360, 2026-07-21
Source: @giaki3003, 2026-07-21
Source: @meatsus, 2026-07-21

40 GB is the stable unlock (August 2026)

On 2026-08-06 arabus reported both of his 10 GB cards unlocked "perfect" as 40 GB cards, and MrManny confirmed to aaasss that 10 GB to 40 GB is the stable unlock; Xeax said people were still hoping for a stable 80 GB. On 2026-08-13 Lycorascence's new 10 GB card passed a five-minute Vulkan memtest at 40 GB once he fixed his cabling (a secondary CPU line from the power supply kept crashing the driver; a Y splitter and a new riser fixed it). The same day cuddylac997 said that a model going over the cap of about 37 GB crashes one card and needs a restart, and that the cmpunlocker2 fork avoids the restart. machinezero referred to the "10-40gb gang" on 2026-08-18. On 2026-08-20 Lycorascence reported that exhausting all VRAM stalls his 40 GB card until the whole machine is rebooted (the reserved top of VRAM: Xid 31 and the reserved top of VRAM), and on 2026-08-23 cuddylac997's leaderboard for four 10 GB cards showed 33.0 GiB verified clean per card. How to unlock and check: Memory unlock on the 10 GB card, Verifying an unlock.

Proven 10 GB
Source: @arabus_74064, 2026-08-06
Source: @aaasss1456, 2026-08-06
Source: @mrmannyatx, 2026-08-06
Source: @xeax, 2026-08-06
Source: @lycorascence, 2026-08-13
Source: @cuddylac997, 2026-08-13
Source: @machinezer0, 2026-08-18
Source: @lycorascence, 2026-08-20

"40 usable, 80 detectable": the working figure from July to September

Through late July and August the 10 GB card was a 40 GB card in everyday talk. On 2026-07-20 meatsus asked Petri Krohn whether owners should "stick to the 40GB unlock", and on 2026-07-22 Luna called 64/40 GB the stable configurations. Eksitus's card reported 40960MiB in nvidia-smi at a 250 W limit and 42 W in P0 at 38 C idle (2026-07-22). DaLZy (2026-07-27) had seen no 10 GB card that failed to unlock to 40 GB. On 2026-07-30 gpulover and Arcanum told a prospective buyer that the card unlocks to 40 GB usable and that touching the other 40 GB crashes; Moliminous added that people could load the full 80 GB but it crashed once errors accumulated, and gpulover advised against paying $850 for a 10 GB card. CUTE (2026-07-31): 10 GB cards unlock to 40 GB, and 80 GB seemed possible but did not work yet. On 2026-08-08 meatsus said nobody had a stable 80 GB unlock, and Humvee55 said that work was shelved while the community worked on PCIe Gen3. Buyers priced the card accordingly: machinezero found 40 GB too little for a four-card DeepSeek setup at full weights and put a fair price at $700 at most for a tested card, and B16_c (2026-08-20) chose 8 GB cards so as not to "be stuck at 40gb". On 2026-08-25 ST hoped "that 80gb unlock" would not happen soon, so none was public; RedPandaMining's six 10 GB cards ran at 40 GB, benchmarked by iszo as a "6x 40gb rig" (2026-08-26); and on 2026-08-30 ST told cuddylac997 that neither PCIe 4.0 nor 80 GB had been unlocked on the 10 GB card. On 2026-09-12 Far answered a newcomer with "40 usable, 80 detectable", and gnasherrr recalled that after the first Reddit post everyone thought the 10 GB card was the one to buy, but the 8 GB card turned out to unlock larger.

Reported 10 GB
Source: @meatsus, 2026-07-20
Source: @lunaroseoc, 2026-07-20
Source: @eksitus, 2026-07-22
Source: @dalzyyy, 2026-07-27
Source: @meromgreen, 2026-07-27
Source: @danske1, 2026-07-27
Source: @ju5t5omedude, 2026-07-27
Source: @artistosu, 2026-07-27
Source: @artistosu, 2026-07-30
Source: @_arcanum_, 2026-07-30
Source: @moliminous, 2026-07-30
Source: @dalzyyy, 2026-07-31
Source: @cute_sc2, 2026-07-31
Source: @sandberg54, 2026-07-31
Source: @roger_dammit, 2026-08-01
Source: @moliminous, 2026-08-01
Source: @lunaroseoc, 2026-08-01
Source: @mv.nyc, 2026-08-06
Source: @meatsus, 2026-08-06
Source: @scallion_9883, 2026-08-06
Source: @pasteyy., 2026-08-06
Source: @cuddylac997, 2026-08-10
Source: @giaki3003, 2026-08-10
Source: @machinezer0, 2026-08-10
Source: @lycorascence, 2026-08-19
Source: @cmurrayis, 2026-08-20
Source: @schakaqt, 2026-08-20
Source: @skirtchaser, 2026-08-20
Source: @redpandamining, 2026-08-21
Source: @lycorascence, 2026-08-21
Source: @stcord, 2026-08-25
Source: @redpandamining, 2026-08-26
Source: @ianszo21, 2026-08-26
Source: @jovial_dolphin_29796, 2026-08-26
Source: @cuddylac997, 2026-08-30
Source: @stcord, 2026-08-30
Source: @giaki3003, 2026-08-30
Source: @irythros, 2026-08-31
Source: @scallion_9883, 2026-09-09
Source: @_arcanum_, 2026-09-09
Source: @sojun80, 2026-09-09
Source: @farawaythrower, 2026-09-11
Source: @supermanisdum, 2026-09-11
Source: @gnasherrr, 2026-09-11

The 80 GB branch and why it fails

In date order. The short version: the 80 GB geometry makes 80 GB visible, and on-card kernels can write and read much of it, but loading, copying or freeing data above about 40 GB crashes the program, the driver or the GSP firmware. No public build had fixed that by 2026-10-03.

Branch 80 (2026-07-19) and the first debugging

At 05:11 UTC on 2026-07-19 Amogh told 10 GB owners: "I just pushed an update to the branch `80` which *should* enable 80GB of VRAM instead of the usual 40GB." That evening a thread, "Far's 80GB VRAM issue", asked Far for the output of sudo dmesg | grep -iE 'SEC2_DEBUG' after re-adding the 80 GB patches, and Amogh said it "might take a *little* more debugging than I thought". The result stayed unresolved through this history.

Reported 10 GB
Source: @_j3ns, 2026-07-19
Source: @_j3ns, 2026-07-19
Source: @farawaythrower, 2026-07-19

What 80 GB would take: SMD parts, a VBIOS patch, or software alone (contested)

Accounts in the order posted:

• 2026-07-19: @_j3ns said 80 GB on the 10 GB card needs SMD components soldered on ("an absolute pain") and pushed the branch for testing. Source: @_j3ns, 2026-07-19
• 2026-07-19: @quark2world relayed a rumour that software alone unlocks 80 GB but unstably, with VRM parts added to stabilise it; @coolgod_72592 had seen Chinese photos and videos of 80 GB with no mention of hardware changes. Source: @quark2world, 2026-07-19
  Source: @coolgod_72592, 2026-07-19
• 2026-08-10: Amogh said 80 GB on the 10 GB card is a separate problem that needs a VBIOS patch, not only a memory geometry change. Source: @_j3ns, 2026-08-10

Contested 10 GB

LMR value: 0x0000028A for 40 GB, 0x0000028B for 80 GB; neither fixes the crash

On 2026-07-19 197something noted that the update to the 80 branch had corrected the LMR value from 28a (40 GB) to 28b (80 GB). On 2026-07-20 Eksitus found a mismatch: a commit to common/constants.yaml changed it to 0x0000028B while one driver patch still used 0x0000028A. Amogh confirmed 0x0000028B and told him to run sudo ./remove.sh --yes, reboot, then repatch. lxwx reported that building with 28b "doesnt help much" with the crash above 40 GB, so the value alone is not the fix.

Proven 10 GB
Source: @197something, 2026-07-19
Source: @eksitus, 2026-07-20
Source: @_j3ns, 2026-07-20
Source: @lxtend, 2026-07-20

cuda_memtest on the 80 branch runs at 39 GB and fails at 40 GB; power is not the limit

Far tested the 80 branch on 2026-07-19 with open kernel module 610.43.03: cuda_memtest reported a global memory size of 85,545,582,592 bytes but hung after attaching to the device, ran only when allocating 39 GB, and LLM loads hung after about 20 GB. lxwx reported that any write above 40 GB crashes the card at once, and Eksitus found use under 40 GB stable, while going over it crashes the running process and usually needs a reboot. Amogh called 80 GB appearing in nvidia-smi without a rejected boot a success for now. Power is not the limit: a 100 W power limit first let Far's cuda_memtest run over all 80 GB, but it hung again on retry, and the limit did not help Eksitus's LLM load, during which his cards drew no more than about 80 W (2026-07-20). Eksitus also saw a host-memory registration fail (cudaHostRegister ... failed: unknown error) while loading GLM-5.2 Q4 with partial offload on three 10 GB cards at 80 GB, and needed a full reboot before the model loaded with about 32 GB per card.

Proven 10 GB
Source: @farawaythrower, 2026-07-19
Source: @_j3ns, 2026-07-19
Source: @eksitus, 2026-07-20
Source: @farawaythrower, 2026-07-20
Source: @eksitus, 2026-07-20
Source: @lxtend, 2026-07-20
Source: @eksitus, 2026-07-20

Theory: HBM charge leakage that rises with temperature

Peocua proposed on 2026-07-20 that the 80 GB errors may be charge leakage in the HBM cells that worsens with temperature, and that keeping the HBM hotspot under 60 C might help. She cited an MDPI paper saying leakage current roughly doubles for every 10 C rise, and Jon Pry's paper ("A Canary in the Crypto Mine"), which suggests that a higher refresh rate resolves the issue. Far was sceptical, since the rest of the HBM2 appears stable at 80 C and above, but agreed to check temperatures on his next test.

Theory 10 GB
Source: @giaki3003, 2026-07-19
Source: @yuiayanami, 2026-07-20
Source: @farawaythrower, 2026-07-20

Errors at default timings, wrong unlock settings or a driver problem (contested)

Readings of the crash above 40 GB on 2026-07-26, in the order posted:

• @never_fly_away, citing Jon Pry's paper: the HBM errors at the hardware level, because at default timings the cells do not hold their values long enough. Such errors rarely crash the GPU or the system; they return wrong results now and then and the application relying on them crashes, which matches what people saw. A timing change might remove them at a compute cost (32% for the card in the paper, 30 to 40% by his estimate). His advice: manage expectations with the 10 GB card; the 8 GB card is the better choice. Source: @never_fly_away, 2026-07-26
• @meatsus: the unlock may be applying wrong or too aggressive clocks or timings; some 10 GB cards may run stable at 80 GB, with no guarantee; missing VRM or other board components are not the cause. Source: @meatsus, 2026-07-26
• @btakk_ak47: more likely a driver problem or a wrong unlock method, because only the applications crash while the card stays detected and the system stays up. Source: @btakk_ak47, 2026-07-26

Contested 10 GB

Theory: looser HBM timings or more refresh could stabilise 80 GB

Archimedes suggested on 2026-07-20 that making the Samsung memory stable at 80 GB might mean raising its clock from 1215 MHz to the A100's 1512 MHz rather than lowering it. On 2026-08-06 Dusty44 opened the feature request "Unlock 80GB stable ram on 10G version" (26 reactions): "if clocks need to slow for memory to stabilise so be it". mim answered that the card seems unable to run any clock but its default (the same day asm64 said his driver build could overclock the memory), and on 2026-08-09 Beam said the original paper suggests it should be possible. On 2026-08-29 jiacheng_liu proposed trading looser HBM timings for stability, by analogy with DRAM overclocking, where "maxmem" disables part of memory to stabilise it: staying at 40 GB is effectively maxmem, and even a 50% bandwidth cut at 80 GB would still beat a Mac Studio or a DGX. Writing timings needs privileged Falcon high-security code. xanthe_ said loosening refresh is what Jon Pry's papers name for a stable 80 GB, and Lycorascence said current unlocks are limited by register issues.

Theory 10 GB
Source: @archi.medes, 2026-07-20
Source: @dusty.44, 2026-08-06
Source: @mim04970, 2026-08-06
Source: @beamreaching1, 2026-08-09
Source: @jiacheng_liu, 2026-08-29
Source: @xanthefin, 2026-08-29
Source: @lycorascence, 2026-08-29

Slower memory or more frequent refresh as a route to 80 GB (contested)

Accounts in the order posted:

• 2026-07-24: @samuel_hayden. said 80 GB could be reached by downclocking; @btakk_ak47 noted this was unverified and that no way to lower the memory frequency at driver level had been found, and asked whether the VBIOS controls it; @ju5t5omedude said his own mention of lower clocks was only speculation that the memory was binned for missing its rated speed; @meatsus agreed in principle that lower clocks and higher latency would give stability if the memory cannot handle the clocks. Nobody posted measurements. Source: @samuel_hayden., 2026-07-24
  Source: @btakk_ak47, 2026-07-24
  Source: @ju5t5omedude, 2026-07-24
  Source: @meatsus, 2026-07-24
  Source: @roger_dammit, 2026-07-24
• 2026-09-29: @okrasit. quoted a research document: the full 80 GB is physically present (hash write and read-back) but shows 2,796 errors under stress at the stock refresh interval while running cleanly at 40 GB, and refreshing more often gives 0 errors at about 32% throughput cost. @cute_sc2 said they had tested lowering the HBM speed and it did not fix the folding; @sojun80 argued that 80 GB stable at roughly 33% slower should clearly be pursued. Source: @okrasit., 2026-09-29
  Source: @cute_sc2, 2026-09-29
  Source: @lycorascence, 2026-09-29
  Source: @sojun80, 2026-09-29

Contested 10 GB

Downclocking does not fix the memory errors (2026-08-30)

CUTE said the card can be downclocked with 170tune, but this barely changes idle power, and that downclocking does not fix the memory errors that keep the 80 GB unlock from being stable. HumanDrone8721 suggested the clock control only affects the GPU core clock, and that the memory would need timing and wait-state changes in the memory controller's registers.

Reported 10 GB
Source: @cute_sc2, 2026-08-30
Source: @humandrone8721_60165, 2026-08-30

Would fitting the missing VRM parts stabilise 80 GB? Consensus no (contested)

The exchange of 2026-07-24, in the order posted:

• @tugoubxs speculated that fitting the unpopulated inductors, DrMOS stages and capacitors around the GPU (shown in Matty Ice's PDF) would supply more power and stabilise 80 GB, and @icematty said his team had hoped so too. tugoubxs later added that whether added phases are even used depends on the PWM controller and its configuration. Source: @tugoubxs, 2026-07-24
  Source: @icematty, 2026-07-24
• @meatsus disagreed: the GPU does not sense the VRM phase count, the existing VRM delivers the roughly 250 W full-load draw, and extra phases would only run cooler, so a stable core voltage cannot be made "more stable". @sealclubber and meatsus put the instability down to silicon that did not meet the A100 standard, and @serplay228 said the stacks are defective, since Samsung sold them as 8 GB parts. Asked by @m_3po whether the unlock needs VRM parts, SealClubber said no. Source: @meatsus, 2026-07-24
  Source: @sealclubber, 2026-07-24
  Source: @serplay228, 2026-07-24
  Source: @m_3po, 2026-07-24

Contested 10 GB

Was the paper's 80 GB result reproducible? (July 2026)

ti asked on 2026-07-24 whether a 10 GB card at 80 GB freezes as soon as use passes 40 GB, which would make the paper's 80 GB results impossible to reproduce, and whether those results are genuine. W.H. believed a working method exists but that the paper's author never released it. Studebaker answered the "when will the 80G method be released" question with a video file dated 2026-06-29 and "Chop chop"; the video is not described in text. tugoubxs noted that no one had claimed a stable 80 GB unlock.

Theory 10 GB
Source: @jovial_dolphin_29796, 2026-07-24
Source: @winghorizon., 2026-07-24
Source: @_studebaker_, 2026-07-24
Source: @tugoubxs, 2026-07-24

Addressing and sub-variant theories (July 2026)

W.H. observed that the 10 GB card's VRAM addressing appears to start at 1G and end at 2 rather than at 0, and wondered whether this relates to why only 40 GB is usable. gk_hun suggested that there are 10 GB sub-variants with differently packed HBM, which could explain inconsistent 80 GB behaviour. Both are untested.

Theory 10 GB
Source: @winghorizon., 2026-07-24
Source: @gk_hun, 2026-07-24

Theory: the upper 40 GB is untrained, and the HBM controller takes commands only at boot

On 2026-07-24 Eksitus said he was ordering the x16 capacitor-mod parts, plus whatever else differs from an A100, to help test an 80 GB unlock, and on 2026-07-26 gk_hun said he was busy with the 80 GB unlock. gk_hun's view (2026-07-27): the card boots fully with 80 GB, but "reported 80G /= reliable 80G yet"; the extra memory "is there and can be written single-context, but untrained and without timings", so he asked for the HBM self-report (IEEE 1500) plus the training-result and mode registers. He floated swapping a resistor on a 10 GB card and flashing the 8 GB card's VBIOS so that the memory might train with the Hynix timings. Samuel Hayden had not looked into die-level hardware attacks because their feasibility is low; from his reading the HBM controller only accepts commands from the Falcon core at boot, which is not easy to interrupt or inject into.

Theory 10 GB
Source: @eksitus, 2026-07-24
Source: @gk_hun, 2026-07-26
Source: @samuel_hayden., 2026-07-27

Strap resistor swaps on a 10 GB card do not change its memory size

On 2026-06-23 jiacheng_liu tried every combination of the last three strap resistors on a 10 GB card running the stock BIOS (an AI assistant had suggested the settings to Amogh) and also removed three straps: none changed the reported 10 GB, and one combination kept the system from booting. GMZombie007 said one of his own 10 GB cards runs a strap setting different from its original and that it changed nothing; meatsus recalled earlier strap tests with no result, and Amogh called it a useful debunk. Background on straps and the VBIOS: VBIOS and firmware.

Proven 10 GB
Source: @_j3ns, 2026-06-23
Source: @meatsus, 2026-06-23
Source: @gmzombie, 2026-06-23
Source: @jiacheng_liu, 2026-06-23

No quick path: the HBM control registers are out of the exploit's reach (2026-07-26)

Samuel Hayden said the HBM control registers cannot be reached with the exploit as it stood, so there was likely no quick path to 80 GB on the 10 GB card, and the same applied to ECC. His advice: anyone buying a 10 GB card should not expect more than 40 GB any time soon.

Reported 10 GB
Source: @samuel_hayden., 2026-07-26

Jon Pry's 80 GB method differs from cmpunlocker (contested)

In the order posted:

• 2026-07-25: @kurj86 relayed that, per Jon Pry's paper, he got his card stable at 80 GB with roughly a 35% throughput penalty, likely with silicon-lottery variation; @samuel_hayden. confirmed that Pry did not publish the Falcon emulator needed to map the registers. Source: @kurj86, 2026-07-25
  Source: @samuel_hayden., 2026-07-25
• 2026-08-12: @toffles3179 argued that the community's 80 GB efforts follow the cmpunlocker path, while Jon Pry described a driverless booter feeding an unmodified stock driver, quite different from how the unlock works now; the community may have found an easy 64/40 GB path that can never reach 80 GB, and Jon Pry states publicly that the 10 GB CMP is an 80 GB card. @giaki3003 mocked the idea of "generating a new stack from thin air". Source: @toffles3179, 2026-08-12
  Source: @giaki3003, 2026-08-12

Contested 10 GB

The public 80 GB unlock: the GSP crashes about 10 seconds after the workload exits

ti tested the public 10 GB to 80 GB unlock (the 580 and 610 driver methods) on 2026-07-27: 80 GB is reported correctly, a 39.7 GB gpu-burn passes, and 1 MB written and read around the 50 GB address shows no immediate errors, but about 10 seconds after the program exits the GSP crashes and the GPU needs a reset. After a function-level reset the memory configuration returns to the native 10 GB values, while the paper says its capacity override survives that reset. The paper applies the override with a driverless loader first, resets the GPU, then loads a clean stock driver on an already unlocked GPU; the public methods apply it during driver and GSP startup. His hypothesis, stated as unproven: the community method changes the memory geometry too late, so early firmware state is still 10 GB while later state is 80 GB, an internal mismatch.

Reported 10 GB
Source: @jovial_dolphin_29796, 2026-07-27

80 GB visible, anything above about 40 GB crashes: the consensus by August

Eksitus (three 10 GB cards, 2026-07-27): values load into the upper region and usage looks right, but any attempt to use that memory crashes the program, and the card needs a reboot to be stable again. Asked that day by gk_hun whether anyone was working on a stable 80 GB unlock, chris.via.egnatia answered "no"; Samuel Hayden and K2V agreed the patch unlocks 80 GB but is not stable, and Huzifei (2026-07-28) said his card "isn't even stable at 41GB". xnupanic76 (2026-07-31) ran a 10 GB card at 40 GB without trouble, but "When I unlocked to 80GB, I got write errors from the driver." In the same exchange MrLoki89 found prefill slow, and Arcanum and chleba suggested a larger prompt batch and llama.cpp or vLLM. By 2026-08-10 this was the consensus: Fats wrote "You can get 80GB to show up in nvidia-smi easily but loading more than 40GB will cause a crash", DaLZy called it the "80gb showing up but only 40gb addressed trap we've tripped like 500 times", formatme noted memory tests run "all the way to 79GB" without the memory being usable, Capone said cuddylac997's 80 GB results were "the same result everyone else has gotten", and gpulover said "only 40 is possible". On the 8 GB card the 64 GB unlock is fully usable: DeeJay uses all 64 GB of each card for DeepSeek.

Proven 10 GB
Source: @eksitus, 2026-07-27
Source: @gk_hun, 2026-07-27
Source: @chris.via.egnatia, 2026-07-27
Source: @samuel_hayden., 2026-07-27
Source: @k2v0008, 2026-07-27
Source: @huzifei, 2026-07-28
Source: @xnupanic76, 2026-07-31
Source: @serplay228, 2026-07-31
Source: @_arcanum_, 2026-07-31
Source: @chleba7742, 2026-08-01
Source: @dalzyyy, 2026-08-09
Source: @beefbonecapone, 2026-08-10
Source: @c.03937, 2026-08-10
Source: @deejay9983, 2026-08-10
Source: @counterstrikeaimgod, 2026-08-10
Source: @artistosu, 2026-08-10

AI coding agents did not crack the 80 GB unlock (July 2026)

formatme ran a coding agent for days trying to unlock 80 GB (GLM 5.2, then GPT 5.6 sol) with no success ("ran a coding agent for days trying to unlock 80GB, nothing"), keeping a progress file and agent memory, with one qemu per GPU. York Zhang said Claude is essential for the reverse engineering. The consensus, with Studebaker, was that real progress needs structured reverse engineering rather than prompting alone, plus many agents and money. How to test a claimed result: Verifying an unlock.

Reported 10 GB
Source: @counterstrikeaimgod, 2026-07-25
Source: @yorkzhang_37447, 2026-07-25
Source: @_studebaker_, 2026-07-25

The 80 branch does not work (2026-08-06): clocks or L2 folding? (contested)

Views on 2026-08-06, in the order posted:

• @c133: "the `80` branch of cmpunlocker doesn't work", do not use it; @ju5t5omedude: "No one has confirmed a properly working 80gb unlock". Only one person, referred to as "Mr John", had claimed a working unlock, without sharing details. Source: @c133, 2026-08-06
  Source: @ju5t5omedude, 2026-08-06
• clee asked about a "driverless refire chain" mentioned on the Frontier-80GB page of the Consensus-Protocol/cmp170hx wiki, and @thanos8912 said it came from another Discord server. Source: @thanos8912, 2026-08-06
  Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @c133, 2026-08-06
• @giaki3003 thought the clocks were not sticking; @zorg33 said it has nothing to do with clocks: "L2 geometry folds 2*40GB and does not address actual 80GB", and it is not certain the upper 40 GB is physically present. Source: @giaki3003, 2026-08-06
  Source: @zorg33, 2026-08-06

Contested 10 GB

The binning argument: an 80 GB-stable die would have been sold as an A100

Jeremy (Ends) argued on 2026-08-06 that any die with enough working HBM for a stable 80 GB would have been sold as an A100 80GB, worth about four times as much, so stable 80 GB 170HX cards are statistically unlikely apart from marginal dies that failed binning; he estimated an HBM failure rate of 10 to 20%. prometasfw countered that a die can have good stacks but failed SMs, full memory without full compute, which could not become an A100, and Humvee55 confirmed he has a 10 GB card with only 56 active SMs; Jeremy agreed this may be the key to finding an 80 GB-stable card. LetterC67 suggested that demand for the 170HX may have caused down-binning.

Theory 10 GB
Source: @endsjeremy, 2026-08-06
Source: @prometasfw, 2026-08-06
Source: @scallion_9883, 2026-08-06
Source: @ju5t5omedude, 2026-08-06
Source: @letterc67, 2026-08-06

80 GB reported in the clean-room server (August 2026)

Rubbra_F said on 2026-08-09 that someone got 80 GB on the 10 GB card in the "clean room", a separate Discord server, and ju5t5omedude said similar work had been done there about a month earlier (around early July 2026). DaLZy named the clean-room server and Pidan Xiong (the kkk.rs author) as two independent confirmations of half-enabled stacks, and clee linked another clean-room message from the same day. The clean-room posts are in another server and cannot be checked from this server's history.

Reported 10 GB
Source: @rubbra_f, 2026-08-09
Source: @ju5t5omedude, 2026-08-09
Source: @dalzyyy, 2026-08-09
Source: @c133, 2026-08-09

Address wraparound: writes above 40 GB silently alias into the first 40 GB

lxwx warned on 2026-08-10 that on a 10 GB card configured past 40 GB he can write to the 41 to 80 GB address range without errors, but those writes change the contents of the 1 to 40 GB region. To test a memory expansion honestly, write a pattern that depends on the address, then read every region back and verify; filling with identical data and reading it back proves nothing about capacity (Verifying an unlock).

Reported 10 GB
Source: @lxtend, 2026-08-10

The 40G fold wall (August 2026)

Several analyses converged in August 2026: on the 80 GB geometry, every address above 40 GB folds back onto the low region that holds the GSP's working memory, so writes up there land on memory the firmware needs and the card crashes with GSP timeouts or an Xid. Changing the memory configuration alone makes the card report 80 GB while the upper 40 GB aliases the lower 40 GB, and the kkk.rs write-up showed the same aliasing (#Reading guides from China). SK Hynix 8 GB cards reach 64 GB with the configuration change alone. icyoung7 says at least two teams reproduced the folding.

Pointer: a longer analysis by @icyoung77666 (the root cause, his reading of the Samsung dies and the routes that remain) exists; see Discord. It waits for the author. The verdict of his public repository is summarised under #Icyoung/cmpunlocker: the 40G fold wall verdict.

Reported 10 GB
Source: @meromgreen, 2026-08-11
Source: @schakaqt, 2026-08-11
Source: @artistosu, 2026-08-11
Source: @icyoung77666, 2026-08-12

Petri Krohn disputes the folding reading (contested)

The fold-wall reading above, against this, in the order posted:

• 2026-08-12: @petrikrohn_26497 rejected the folding explanation: aliasing just means the upper memory is not there (or address lines are missing), and the real 80 GB problem is getting CUDA to use the full 80 GB. His test: apply the unlock with the driverless unlocker, do not load the driver, then write and read back more than 40 GB of unique data (hashes or per-slot addresses, not ones and zeros) through the frame buffer. He says others did this successfully and that it is the basis of the 80 GB physical-memory claim. He warns that the GPU cache can return fake values for the upper half if not all 80 GB is written, and that after the configuration writes a card may simply advertise 80 GB that does not exist. Source: @petrikrohn_26497, 2026-08-12
• 2026-08-12: @meatsus warned that leading questions to an LLM ("can the addresses fold?") poison its context. Source: @meatsus, 2026-08-12

Contested 10 GB

Competing theories of what blocks 80 GB (August 2026)

AKuchen held that the limit sits in the HBM logic die rather than in the fuses, that the Chinese 80 GB group was stuck on the logic die's IEEE 1500 configuration, and that a JTAG chain reaching the HBM test interface could be a way in without a Samsung datasheet; aaasss thought an A100 80 GB would be needed as a reference. Big Ptoughneigh was probing the HBM test interface and modifying the GSP and driver, and argued that a device-ID mismatch could let the PMU derive the extra address bit needed for 80 GB; asked for proof of 16 Gb dies, he gave none. cuddylac997 blamed a GSP DMA-tracking table, and Zarknoth saw the GSP hang in the free path above 48 GB. None of these was proven.

Theory 10 GB
Source: @akuchen_66498, 2026-08-09
Source: @aaasss1456, 2026-08-09
Source: @cuddylac997, 2026-08-10
Source: @zarknoth, 2026-08-10
Source: @bigptoughneigh, 2026-08-12

Zarknoth: 79 GB of writes in 128 MB chunks, then a crash on teardown

Working with Codex, Zarknoth got 79 GB of full writes, but the card crashed on teardown and driver reset because CUDA could not free the memory. Chunking allocations into 128 MB pieces allowed allocation above 40 GB but stopped at 48 GB unless the free path was fixed to stop the GSP hanging above 48 GB; he split the memory allocator into at least six address windows, and the 50 to 64 GB window failed most. A later test with a different memory configuration showed 80 GB, but allocation ran out of memory at 39 GB with illegal-address errors above it; lxwx suggested the 80 GB LMR value, then said it changes nothing. Zarknoth went back to 40 GB.

Reported 10 GB
Source: @zarknoth, 2026-08-10
Source: @lxtend, 2026-08-10
Source: @hastwell, 2026-08-10

A claimed 76 GB on four cards (contested)

In the order posted:

• 2026-08-10: @cuddylac997 posted a test log allocating 0.25 GiB chunks up to 75.00 GiB with fill and verify, claimed 76 GB visible in nvidia-smi and verified on four cards, and said the remaining block is a firmware lock in the GSP's DMA-tracking table. Source: @cuddylac997, 2026-08-10
• 2026-08-10: @beefbonecapone noted that his earlier output said DMA did not work above 40 GB and that every screenshot showed Xid crashes; @giaki3003 called it fake. @toffles3179 argued that a 40 GB physical limit would make Jon Pry wrong, and @meatsus noted that Jon Pry had changed timings and memory clocks. A verify pass has to use address-dependent patterns because of the wraparound. Source: @beefbonecapone, 2026-08-10
  Source: @giaki3003, 2026-08-10
  Source: @toffles3179, 2026-08-10
  Source: @meatsus, 2026-08-10
• 2026-09-30: cuddylac997 said that when he first joined he had reported 76 GB showing on a 10 GB card and was talked out of continuing, and that he has a warehouse full of cards. @_____josh_____ suggested he wants people to brick cards so he can sell them; @con3po7016 vouched that there is enough evidence he is legit. Source: @cuddylac997, 2026-09-30
  Source: @con3po7016, 2026-09-30
  Source: @_____josh_____, 2026-09-30

His later gist and its retraction are under #cuddylac997's gist: DMA above about 40 GiB crashes the GSP (contested).

Contested 10 GB

Firmware lock or binned memory? The August dispute over the upper 40 GB (contested)

In the order posted:

• 2026-08-10: Massimo confirmed the 40 GB unlock on his new 10 GB card. @giaki3003 said the 10 GB cards in circulation have only 40 GB physically (an 80 GB-capable one may exist, but he had never seen one); @mv.nyc and @xeax said the card has 80 GB physically with only 40 GB stable, Xeax adding that one person claims stability above 40 GB by giving up about 30% memory throughput and that most cards are binned, so results vary. @redpandamining had never seen a 10 GB card sustained at 80 GB, giaki3003 called the consensus that 80 GB is "not a thing" on these cards, and @port00 relayed a claim from another chat that higher temperatures let a card pass its 80 GB test. Source: @mv.nyc, 2026-08-10
  Source: @giaki3003, 2026-08-10
  Source: @xeax, 2026-08-10
  Source: @celex3735, 2026-08-10
  Source: @redpandamining, 2026-08-10
  Source: @giaki3003, 2026-08-10
  Source: @port00, 2026-08-10
• 2026-08-10: giaki3003 told @cuddylac997 that 80 GB on the 10 GB cards is "physically impossible". giaki3003 and @counterstrikeaimgod argued there is no proof the 80 GB is usable: tests passing beyond 40 GB had been known for over a month, and in their view the HBM is binned and stable for half its capacity. formatme pointed to the kkk.rs write-up as proof that the Chinese community had tried harder and failed; giaki3003 read it as showing the 80 GB is not there, while @dalzyyy (who reads Chinese) and Xeax read it as showing that it is. cuddylac997 argued for a firmware limit rather than binning, because every card stops at exactly 40 GB; @mattvirus suggested some extra capacity might work but be walled off; @impactic understood the card as 40 GB or 80 GB, never 60 GB. Source: @cuddylac997, 2026-08-10
  Source: @giaki3003, 2026-08-10
  Source: @machinezer0, 2026-08-10
  Source: @giaki3003, 2026-08-10
  Source: @counterstrikeaimgod, 2026-08-10
  Source: @cuddylac997, 2026-08-10
  Source: @dalzyyy, 2026-08-10
  Source: @mattvirus, 2026-08-10
  Source: @impactic, 2026-08-10
  Source: @xeax, 2026-08-10
• 2026-08-21: @ju5t5omedude said the 10 GB card unlocks to 40 GB only, reliably, for the foreseeable future, and @_studebaker_ summed it up as "10GB if you prefer Samsung over Hynix and playing blackjack". @irythros said 80 GB can be reached but the memory is corrupted (addresses wrap after 40 GB and silently overwrite earlier data), and @schakaqt said, per Chinese articles, that even when 80 GB looks stable the card silently falls back to the first half. @xanthefin countered that Jon Pry's paper describes 80 GB first unstable and then made stable by a tweak at some performance cost. @skurtdiggler, by contrast, bought the 8 GB card and got 64 GB with all memory checking out. Source: @_studebaker_, 2026-08-21
  Source: @ju5t5omedude, 2026-08-21
  Source: @irythros, 2026-08-21
  Source: @schakaqt, 2026-08-21
  Source: @xanthefin, 2026-08-21
  Source: @skurtdiggler, 2026-08-21
  Source: @jo29_29, 2026-08-21
• 2026-08-22: @artistosu said the 10 GB card is worth half price at most; DaLZy said the remaining 40 GB is supposed to be faulty; @meromgreen and @_arcanum_ compared Samsung's and SK Hynix's HBM records; @aaasss1456 replied (in Chinese) that these are guesses he would believe only with X-ray photos. Source: @jo29_29, 2026-08-22
  Source: @artistosu, 2026-08-22
  Source: @dalzyyy, 2026-08-22
  Source: @meromgreen, 2026-08-22
  Source: @_arcanum_, 2026-08-22
  Source: @aaasss1456, 2026-08-22
• 2026-09-13: @cute_sc2 still described the 10 GB card as cheaper "despite the only 40gb unlock", while @djpandaman joked that eBay buyers were fighting because "the 10gb card is cheaper AND you can get 80gb". Source: @cute_sc2, 2026-09-13
  Source: @djpandaman, 2026-09-13

Contested 10 GB

Theory: better 10 GB cards are being picked out by serial number

Studebaker suggested on 2026-08-10 that buyers may be selecting the better 10 GB cards by serial number, as with the CMP 100-210, where the V100-compatible boards were picked out and the rest sold on. No evidence was posted, and no serial ranges are known to map to better cards.

Theory 10 GB
Source: @_studebaker_, 2026-08-10

Capacity ceilings and the "not software-unlockable" argument (contested)

The thread of 2026-08-16, in the order posted:

• @lunaroseoc gave the usable ceilings as 40 or 80 GB for the 10 GB card and 64 or 96 GB for the 8 GB card, saying 80 GB works but only about two fully functional 80 GB cases had been recorded, and that the top one or two stacks are usually defective. Source: @lunaroseoc, 2026-08-16
• @giaki3003 claimed in this thread that the extra memory is only software-deactivated, which differs from his "physically impossible" of 2026-08-10. Source: @giaki3003, 2026-08-16
• @light_2212 argued that without NVIDIA's signing key or fuse authority, linear 80 GB, PCIe 4.0, NVLink and ECC are not software-unlockable, linking the Icyoung/cmpunlocker research; @bayley_w countered that the whole toolchain is exploits anyway. Source: repo github.com/Icyoung/cmpunlocker by Icyoung on GitHub, posted by @light_2212, 2026-08-16
  Source: @bayley_w, 2026-08-16

Contested 8 GB and 10 GB

Status on 2026-08-25: readable and writable, but crashes (contested)

Views on 2026-08-25, in the order posted:

• @_maitops: the 10 GB card has 80 GB of HBM2, proven, but not stable; the memory can be read and written, but using it crashes. Source: @_maitops, 2026-08-25
• @cute_sc2 thought some cards are 40 GB and some 80 GB. Source: @cute_sc2, 2026-08-25
• @ob88594: the key issue is addressing, since the last 40 GB of address space cannot be read or written correctly (citing the Chinese blog post "170Hx 10G解锁80G探索"). Source: @ob88594, 2026-08-25
• @lycorascence: the unlock needs deeper access than driver patching gives. Source: @lycorascence, 2026-08-25

Contested 10 GB

The memory is there and the GSP crashes managing it (MaitOPS, August to September)

MaitOPS, who has three 10 GB cards, reposted cuddylac997's gist (gist) in the 80 GB feature-request channel on 2026-08-30 and on 2026-09-01 said he was starting on the problem. On 2026-09-05 he said he had run experiments using memory above 40 GB on all three cards: "There is 80GB all experiments proved it. It's just not usable today, it will maybe never be, but the HBM is there." The same day regaly explained why the 10 GB to 40 GB unlock works at all: the full GA100 configuration carries 80 GB of HBM, so the unlock exposes memory that exists, and nothing can be unlocked beyond that maximum. On 2026-09-17 MaitOPS summed up that the 80 GB is present and usable by the SMs, but the GSP crashes when managing allocations above 40 GB, so CUDA cannot use the extra memory. On 2026-09-19 he said Jon Pry was the first to find the 80 GB of HBM on the 10 GB card, that the crash hits the same point on several cards, and that his own GLM 5.3-assisted reverse engineering points to a bug in the GSP memory allocation map; the GSP firmware could not be patched at that time. The gist itself is contested (below).

Reported 10 GB
Source: @_maitops, 2026-08-30
Source: @_maitops, 2026-09-01
Source: @_maitops, 2026-09-05
Source: @_maitops, 2026-09-05
Source: @_maitops, 2026-09-05
Source: @_maitops, 2026-09-05
Source: @reggie7756, 2026-09-05
Source: @_maitops, 2026-09-05
Source: @_maitops, 2026-09-17
Source: @_maitops, 2026-09-19
Source: @lycorascence, 2026-09-19

A 76 GiB allocation test threw an Xid (2026-09-04)

In the 80 GB feature-request channel squishysquare said he was testing a 76 GiB CUDA allocation. It threw an Xid error, which he blamed on his test harness rather than on the memory tests themselves. He posted no follow-up result in that channel.

Reported 10 GB
Source: @squishysquare, 2026-09-04

An addressing problem, not stability (September 2026) (contested)

Views on 2026-09-16, in the order posted:

• @cute_sc2: the 80 GB problem is addressing, not stability: accesses above 40 GB point to the L2 cache instead, so either the memory is not there or it is a firmware or configuration issue. Source: @cute_sc2, 2026-09-16
• @lia_oc: the card unlocks to 40 GB; 80 GB is not stable and not expected soon. Source: @lia_oc, 2026-09-16
• @_____josh_____: the stacks are defective (otherwise NVIDIA would have sold them as 80 GB A100s), and bypassing the guardrails would only give errors. Source: @_____josh_____, 2026-09-16
• @dalzyyy: the memory may work at a lower pin speed than NVIDIA required. Source: @dalzyyy, 2026-09-16

Contested 10 GB

cuddylac997's gist: DMA above about 40 GiB crashes the GSP (contested)

In the order posted:

• August 2026: @cuddylac997 (Linux, four 10 GB cards, PCIe 2.0 x4) got nvidia-smi to show 80 GB (about 76 GB usable after driver load), a Claude-written test "verified good to 75.00 GiB" and about 780 GB/s of effective HBM bandwidth, yet nothing loaded above 40 GB (GLM 5.2 failed; DeepSeek V4 Flash ran on four cards only with each card under 40 GB). His gist, "CMP 170HX 80GB geometry: DMA above ~40 GiB crashes GSP", first posted on 2026-08-10, blamed a GSP DMA-tracking table and a two-phase lock. Source: @cuddylac997, 2026-08-08
• 2026-09-30: with the 80 GB geometry applied, his hbm45 CUDA-kernel test, which writes a unique pattern to every chunk and verifies only after all writes, came back clean to at least 76 GiB on three 10 GB cards with no aliasing, while host-to-device copies above about 40 GiB fault. In his reading 80 GB is physically real but needs a GSP DMA patch, or frameworks that load weights through kernel writes. @jacksmith7981 summarised the updated gist (LLM-assisted, built on his own test report). This contradicts the folding reports; @sojun80 countered that the memory is accessible but not stable, and @lycorascence suggested card-critical memory may land in writable memory unless that range is cordoned off. @_orrorin and Lycorascence volunteered to test. Source: @cuddylac997, 2026-09-30
  Source: @jacksmith7981, 2026-09-30
  Source: @lycorascence, 2026-09-30
  Source: @sojun80, 2026-09-30
  Source: @_orrorin, 2026-09-30
  Source: @jacksmith7981, 2026-09-30
• 2026-09-30, corrections the same day: the trigger is copy-engine DMA once total card allocation passes about 38.5 to 44 GiB, not the destination address, and runs with a single GPU or a small KV cache still fault. The claim that a staging buffer (copy low, then move the data up with a kernel) bypasses it was wrong, as JackSmith showed from the gist's own table, and was retracted. Below the line the card works (DeepSeek-V4-Flash at 38.5 GiB per card, 306.8 t/s prefill), and "there is no proven software workaround today". Source: @cuddylac997, 2026-09-30
• 2026-09-30: reproduction failed. Lycorascence, with the 80 GB geometry applied (nvtop showed 80 GB), crashed the driver and hung the PC at about 39.9 GiB with hbm45, and cuddylac997's own later test crashed a card above 40 GB. Source: @lycorascence, 2026-09-30
• 2026-10-01: cuddylac997 wrote "A lot of it was wrong. Making a new one" and took the repository down. Source: @cuddylac997, 2026-10-01
• 2026-10-03: @deaddudehangin could not get around the folding: his writes always landed in the lower 40 GB. Source: @deaddudehangin, 2026-10-03

Contested 10 GB

Allocating 45 GiB is safe; writing or freeing it faults

In an AI-run test on the 80 GB geometry (cuddylac997, 2026-09-30), one fixed 512 MiB buffer was written and verified while idle cudaMalloc padding was raised to 36, 38, 40, 42 and 45 GiB; holding 45 GiB of untouched allocations stayed clean, and only the teardown faulted. Reserving the memory is not the trigger; writing, loading or freeing it is. Lycorascence likewise saw the map command reserve the full 80 GB.

Reported 10 GB
Source: @cuddylac997, 2026-09-30
Source: @lycorascence, 2026-10-01

ggualerz/cmp170hx10g-to-80g: 80 GB not achieved

A research repository of Markdown reports (the source, patches and tools are not public) on taking the 10 GB card from 40 GB to 80 GB with a cmpunlocker-patched driver. Its conclusion: 80 GB is not achieved. With a coherent 80 GB configuration nvidia-smi reports 81920 MiB and an SM-kernel fill and verify of 79 GiB ran clean, but every host-to-device write above 40 GB faulted (Xid 119, 1 and 154). The teardown crash threshold bisected to 158 x 256 MiB (39.50 GiB), and two crash dumps were byte-identical, which points to deterministic software rather than HBM decay. The old 80 branch paired the 80 GB memory configuration with the 40 GB LMR value, an incoherent pair that folds at exactly 40 GiB. The author also found and fixed a late-PMA overlap bug in cmpunlocker and lists dead ends not to retry. zee9 posted it in the server.

Reported 10 GB
Source: repo github.com/ggualerz/cmp170hx10g-to-80g by ggualerz on GitHub, posted by @ziwhon, 2026-09-06

Icyoung/cmpunlocker: the 40G fold wall verdict

A standalone copy of cmpunlocker with an experimental 80 GB profile for the 10 GB card (81920 MiB) and the whole August 2026 research archive (experiments v1 to v64 and 22 notes). Its verdict: linear 80 GB is not software-unlockable because of the "40G fold wall": writes above 40 GB fold back at a +35 GiB offset, leaving a safe pool of about 31 to 36 GiB, and it recommends --profile=10gb (40 GB) for production. The repository traces the difference between the CMP and the A100 to per-partition values in a signed firmware table (the FWSEC devinit table), which a modified flash would likely fail to pass. Its own documents say that earlier "60/67G PASS" records were fold artifacts and that the 80 GB screenshots do not show safely usable memory. Apoloniato, emogop and LightOverDarkness posted it in the server.

Reported 10 GB
Source: repo github.com/Icyoung/cmpunlocker by Icyoung on GitHub, posted by @light_2212, 2026-08-16
Source: repo github.com/Icyoung/cmpunlocker by Icyoung on GitHub

zerobitcopy/cmp-easyunlock: its own README calls the 80 GB mode unstable

A one-script unlock that builds a bundled patched copy of the open driver 580.159.03, swaps in modified GSP firmware and loads it, temporarily or persistently. Its four modes include 10g_to_40g and 10g_to_80g, and the README warns that 10g_to_80g is unstable. tugoubxs asked in the server on 2026-07-24 whether it belonged to the server admin; nobody answered, so its authorship is unknown, and it is not a cmpunlocker fork. adih linked it on 2026-07-28 as a possible source of a Xianyu seller's 80 GB method (#Fake 80 GB listings). Trust: it ships a prebuilt binary and firmware blobs nobody can audit, downloads and installs the NVIDIA driver itself, and reboots without asking in persistent mode. The maintained unlock: Driver unlock.

Reported 10 GB
Source: repo github.com/zerobitcopy/cmp-easyunlock by zerobitcopy on GitHub, posted by @tugoubxs, 2026-07-24
Source: repo github.com/zerobitcopy/cmp-easyunlock by zerobitcopy on GitHub, posted by @adih1266, 2026-07-28

d3dx9/cmpunlocker: an abandoned driverless unlock that defaults to 80 GB

A driverless unlock for the open driver 580.x that installs a service re-applying the volatile unlock at boot. Its default target is an 80 GB profile (5 x 16 GiB), with 40 GB and the native 10 GB as options, and its README asserts that the dies are 16 GB each. Per Petri Krohn's history this is the driverless path that was abandoned; the driver-modifying lineage won out. Amogh linked it on 2026-07-19 to test its PCIe Gen4 claim and reported that it "didn't do anything" for him; Far's board is also Gen3, and running it gave him a black screen, which he blamed on a wrong driver install. Nobody confirmed a speed gain or usable 80 GB from it, and its last commit is from 2026-07-18.

Reported 8 GB and 10 GB
Source: repo github.com/d3dx9/cmpunlocker by d3dx9 on GitHub, posted by @_j3ns, 2026-07-19
Source: @lunaroseoc, 2026-07-19
Source: @farawaythrower, 2026-07-20
Source: @petrikrohn_26497, 2026-08-11

kinako404/cmpunlocker: an 80 GB screenshot, never confirmed

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, and nobody confirmed usable 80 GB from it. The fork could no longer be found when this wiki collected the community repositories in October 2026. Treat its claim as unverified.

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

mintoriakamoto/cmpunlocker2: the "full 80 GB" claim is regarded as fake

Warning 10 GB On 2026-09-10 DeeJay flagged github.com/mintoriakamoto/cmpunlocker2, whose README then claimed PCIe Gen5 x16 on the 170HX and, per clee, the full 80 GB on 10 GB cards. Amogh said Gen5 is not possible on GA100 SKUs, Capone called it "100% slop" and acedogblast agreed it was too good to be true. A later version of the README covers only the 10 GB card at 40 GB and says 80 GB is blocked by firmware. On 2026-10-09 68x said the same developer had claimed PCIe 5.0 on the 170HX in a cmpunlocker issue, and Amogh replied that this was exactly why he had banned them. Do not use it for 80 GB.

Source: repo github.com/mintoriakamoto/cmpunlocker2 by mintoriakamoto on GitHub, posted by @deejay9983, 2026-09-10
Source: @c133, 2026-09-10
Source: @_j3ns, 2026-09-10
Source: @schakaqt, 2026-09-10
Source: @acedogblast, 2026-09-11
Source: @beefbonecapone, 2026-09-11
Source: @68x0, 2026-10-09
Source: @_j3ns, 2026-10-09

48 GB and the sixth stack

A GA100 package has six HBM stack sites, and the 10 GB card uses five. Whether the sixth stack is present, usable or a dummy decides whether 48 GB (6 x 8 GB) is possible. The RIR route that later reached 48 GB on some cards is on RIR fuse burn; reading a card's fuses is on Fuses and floorsweep maps.

The 10 GB card's RIRs were left empty (Humvee55, September 2026)

On 2026-09-29 Humvee55 said NVIDIA apparently forgot to fill the 10 GB card's RIRs with junk, as it did on other GA100 variants. Through the sw-iff route, SMs up to 126, one extra memory stack (up to 48 GB if that stack is not marked defective) and PCIe Gen3 or Gen4 all work; for compute the 10 GB card wins, while for memory the 8 GB card (64 GB) remains king. reala concluded there is little hope for the same unlock on the 8 GB card if it depends on empty RIRs. Earlier, on 2026-09-08, Humvee55 had offered to buy 10 GB cards or trade 8 GB cards for them, saying he needed more SKUs to test on.

Reported 10 GB
Source: @scallion_9883, 2026-09-08
Source: @seriousrising, 2026-09-08
Source: @stcord, 2026-09-08
Source: @scallion_9883, 2026-09-29
Source: @reala7096, 2026-09-29

48 GB claimed on some 10 GB cards: "no more than 48 GB on board" (contested)

In the order posted:

• 2026-09-18: @xtj7 asked whether 10 GB cards with nothing marked defective in gpcprobe could run stable at 80 GB; @scallion_9883 said no: at most one extra stack can be unlocked on a 10 GB card, for 48 GB stable, and only some "unicorn" cards can enable it. Members still ran their cards at 40 GB: @romangitz unlocked one to 40 GB, @squishysquare runs two "40G" cards and @exile_1ck had a "40GB 170hx" on order. Source: @xtj7, 2026-09-18
  Source: @scallion_9883, 2026-09-18
  Source: @romangitz, 2026-09-18
  Source: @squishysquare, 2026-09-18
  Source: @exile_1ck, 2026-09-18
  Source: @okrasit., 2026-09-18
  Source: @scallion_9883, 2026-09-18
• 2026-09-26: @cute_sc2 said 10 GB cards "as of now" carry no more than 48 GB of HBM on board, so the maximum is a 6 x 8 GB configuration, and only if no HBM die is defective; @con3po7016 recalled that the card needed a higher memory refresh rate to reach 80 GB. Source: @cute_sc2, 2026-09-26
  Source: @con3po7016, 2026-09-26

Contested 10 GB

gpcprobe on 10 GB cards: 48 GiB maximum, or "80 GB inside" (contested)

In the order posted (from 2026-09-18):

• Gpcprobe on @_orrorin's 10 GB card showed 48 GiB max (@matstatman: "you should be able to get 8gb more? 40gb -> 48gb"). @lycorascence's card showed six stacks, four full (8 GiB) and two half (4 GiB), 20 of 24 FBPAs = 5120-bit, 40 GiB exposed and 48 GiB max assuming 1 GiB dies. @cute_sc2 guessed that some 10 GB variants unlock to 48 or 96 GB; @jjb01957 thought 48 was the most to hope for.
• @_maitops countered that 80 GB of HBM is inside, readable and writable from the SMs, but normal workloads crash the GSP firmware. Source: @matstatman, 2026-09-18
  Source: @_orrorin, 2026-09-18
  Source: @lycorascence, 2026-09-18
  Source: @cute_sc2, 2026-09-18
  Source: @jjb01957, 2026-09-18
  Source: @_maitops, 2026-09-18

Contested 10 GB

Enabling the sixth stack moves the fold from 40 GB to about 48 GB (contested)

In the order posted (2026-09-29):

• @lycorascence: until folding is fixed, the most a 40 GB card gets is 48 GB, and only if the stacks are not defective. Source: @lycorascence, 2026-09-29
• @cute_sc2: with the original five 8 GB stacks, folding starts at 40 GB; if the disabled stack is enabled (6 x 8 GB) by using a RIR to clear its memory-disable information, folding happens at about 48 GB and all 48 GB is usable. He said this was tested in the developer Discord in late August 2026. Source: @cute_sc2, 2026-09-29
• @djpandaman had seen no evidence that 48 GB is possible, and noted that a 48 GB card depends on the sixth stack, so it cannot then become 80 GB unless that stack is disabled again. Source: @djpandaman, 2026-09-29

Contested 10 GB

Pidan Bear's two 10 GB cards: six stacks and half-enabled stacks (contested)

The thread of 2026-08-09, in the order posted:

• @aaasss1456 relayed that Pidan Bear has two 10 GB cards: on one all six HBM positions are populated, and with the eFuse enabled four stacks are fully used and two partly; the other uses only five stacks. The original report is in another server.
• @bayley_w disputed it: all A100s are built with six stacks and one is disabled after CoWoS assembly for yield; stacks are 1024 bits and the bus 5120 bits, so six active stacks would need two at 512 bits, and the deciding test is HBM2 against HBM2e, since no 16 GB HBM2 stack exists.
• @dalzyyy answered that thanos has a card with half-enabled stacks, that the memory controller can address 512 bits at a time, and that this was confirmed independently twice (by Pidan Xiong and by the clean-room server); he also thinks the 10 GB card's "HBM2" is binned 80 GB-class HBM2e that Samsung sold as 40 GB HBM2. @k2v0008 cited the NVIDIA whitepaper (A100 40G HBM2, A30 HBM2, A100 80G HBM2E, A30X HBM2E) and guessed that all 170HX use HBM2e.
• @_____josh_____ suggested the 10 GB card is capped at 40 GB for stability, bayley_w rejected the idea of building an A100 from five known-bad stacks, and @hastwell suggested damage during assembly. Source: @bayley_w, 2026-08-09
  Source: @aaasss1456, 2026-08-09
  Source: @dalzyyy, 2026-08-09
  Source: @k2v0008, 2026-08-09
  Source: @_____josh_____, 2026-08-09
  Source: @hastwell, 2026-08-09
  Source: @skirtchaser, 2026-08-09

Contested 10 GB

How many stacks sit on the package: six sites, five plus a dummy, or six working stacks? (contested)

In the order posted:

• 2026-07-20: @dot7551 posted a PG199 board with four HBM stacks, and @archi.medes noted a TechPowerUp image of an A100 with six. @_arcanum_: every GA100 package has six stack sites and stacks are disabled per SKU; there is one physical GA100 package, or technically three (Samsung HBM2, Samsung HBM2e and SK Hynix HBM2e). @cute_sc2: the OEM 96 GB A100 is the SKU that uses all six. Source: @dot7551, 2026-07-20
  Source: @archi.medes, 2026-07-20
  Source: @_arcanum_, 2026-07-20
  Source: @cute_sc2, 2026-07-20
  Source: @decoy.works, 2026-07-20
• 2026-07-22: @farawaythrower asked whether two stacks are physically disabled on the 170HX; @giaki3003 said nothing proves it and insider information would be needed, @mrpec said the package has 80 (maybe 96) GB of HBM2e but doubted the exploit can enable it, and @_studebaker_ suggested an X-ray of the package. Source: @farawaythrower, 2026-07-22
  Source: @giaki3003, 2026-07-22
  Source: @mrpec, 2026-07-22
  Source: @jiacheng_liu, 2026-07-22
  Source: @68x0, 2026-07-22
  Source: @_studebaker_, 2026-07-22
• 2026-07-25: @dalzyyy cited a Kurnal Insights GA100 dieshot saying "6 hbm dies (of 2 are dummies)", which would mean some 170HX have only four real stacks; giaki3003 replied that only a side X-ray would show the real count. Source: @dalzyyy, 2026-07-25
  Source: @giaki3003, 2026-07-25
  Source: @winghorizon., 2026-07-25
  Source: @btakk_ak47, 2026-07-25
• 2026-08-06: @xeax said the 10 GB card has five stacks and a wider bus (5 x 16 = 80 GB) and the 8 GB card four (4 x 16 = 64 GB); @turbine believed all cards carry six. @skirtchaser quoted NVIDIA's A100 whitepaper (the full GA100 has six HBM2 stacks and twelve 512-bit memory controllers; no fully enabled A100 ever shipped) and a teardown report (SPR21579) that lists a filler die, meaning five stacks plus one dummy; @robbbbbb7973 uploaded the PDF. Source: @xeax, 2026-08-06
  Source: @turbine, 2026-08-06
  Source: @dalzyyy, 2026-08-06
  Source: @quark2world, 2026-08-06
  Source: @giaki3003, 2026-08-06
  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
• 2026-08-07 and 08-09: SkirtChaser said the failed fifth stack of an 8 GB card is physically removed in production and cannot be recovered; @bayley_w said all A100s are built with six stacks and one is disabled after CoWoS assembly. Source: @skirtchaser, 2026-08-07
  Source: @_studebaker_, 2026-08-07
  Source: @giaki3003, 2026-08-07
  Source: @bayley_w, 2026-08-09
  Source: @aaasss1456, 2026-08-09
  Source: @dalzyyy, 2026-08-09
  Source: @k2v0008, 2026-08-09
  Source: @_____josh_____, 2026-08-09
  Source: @hastwell, 2026-08-09
  Source: @skirtchaser, 2026-08-09
• 2026-08-08 and 08-12: @petrikrohn_26497 disagreed: every 170HX has six working HBM2 stacks, each with two FBPs and four FBPAs; two or four of the FBPs (half-stacks) are fused off as defective or disabled, and every observed card has at least one FBP working on each stack, so there are no dormant stacks: 8 FBPs (8 GB card) or 10 FBPs (10 GB card) spread over six working or half-working stacks. @meatsus said 96 GB would need the full bus and is impossible unless dormant stacks could be unlocked, but also argued that HBM yields are high, so many dies may carry six fully working stacks. Source: @skirtchaser, 2026-08-08
  Source: @petrikrohn_26497, 2026-08-08
  Source: @giaki3003, 2026-08-08
  Source: @quark2world, 2026-08-08
  Source: @meatsus, 2026-08-08
  Source: @petrikrohn_26497, 2026-08-12
  Source: @meatsus, 2026-08-12
  Source: @giaki3003, 2026-08-12
• 2026-08-10: DaLZy cited a launch interview on hardwareluxx saying GA100 packages bond six stacks with one disabled and no dummy dies, so many GA100s might physically carry 96 GB; giaki3003 replied that 96 GB A100s exist but there is no proof CMPs got them, and bayley_w that a 96 GB part turned into 64 GB means the other two stacks are bad. Source: @dalzyyy, 2026-08-10
  Source: @giaki3003, 2026-08-10
  Source: @bayley_w, 2026-08-10
• 2026-08-24: pointed to an A100 96GB listing, @_maitops said it normally uses the same HBM as the 8 GB card and that both cards physically carry 96 GB of HBM2e (six stacks). Source: @confident_lamb_62706, 2026-08-24
  Source: @_maitops, 2026-08-24
• 2026-09-16 and 09-17: @rh3dai argued that six stacks sit physically next to the core, disabled internally by fuses rather than removed; @toffles3179 countered that binning likely happened at interposer bonding (6 stacks = 96 GB, 5 = 80 GB, 4 = 64 GB) and that stacks fused off before NVIDIA took possession may not be bonded correctly. @cute_sc2 said fuse readouts on most 8 GB cards show all six HBM dies with full stacks present, just deactivated; asked by @jjb01957 whether anyone had X-rayed or CT-scanned a card, @wakeupmbrc1 said the six positions are real stacked HBM, 96 GB in total at 16 GB per stack. Source: @rh3dai, 2026-09-16
  Source: @toffles3179, 2026-09-16
  Source: @cute_sc2, 2026-09-16
  Source: @quark2world, 2026-09-16
  Source: @segmond, 2026-09-16
  Source: @jjb01957, 2026-09-17
  Source: @wakeupmbrc1, 2026-09-17
  Source: @cibernox, 2026-09-17

No X-ray or cross-section of a 170HX was posted.

Contested 8 GB and 10 GB

Reading guides from China

kkk.rs: "170Hx 10G unlock 80G exploration"

The most cited write-up on the 80 GB question is a Chinese blog post by the technician 皮蛋熊 (Pidan Bear, also written Pidan Xiong): 170Hx 10G解锁80G探索 (in Chinese). It concludes that there is currently no way to address the upper memory, and describes the block as an HBM or logic-die addressing limitation out of reach. Apoloniato first posted it in the server on 2026-08-09. On 2026-09-30 JackSmith noted that its experiments 2.1 and 2.2 used GSP DMA, so technically they do not contradict the kernel-write result in cuddylac997's gist. Members read its die measurements both ways (Memory unlock on the 10 GB card); read the original before citing it. This wiki links the post and does not translate it.

Reported 10 GB
Source: @apoloniato_07709, 2026-08-09
Source: @jacksmith7981, 2026-09-30

170hx.club (HX Studio)

On 2026-08-12 llm_ftw found, through Bilibili, a Chinese website about unlocking the 10 GB card, 170hx.club ("HX Studio, 170HX AI Video"), and asked Chinese speakers to check it. Its content was not evaluated in this history.

Reported 10 GB
Source: @llm_ftw_11550, 2026-08-12

Fake 80 GB listings

A card can be made to show 80 GB in nvidia-smi without the memory being usable, and a screenshot is easy to edit. Test any card sold as "80 GB" by writing and verifying data above 40 GB (Verifying an unlock).

A Xianyu seller claiming a stable 80 GB (July 2026) (contested)

In the order posted (2026-07-28):

• @jovial_dolphin_29796 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; @sssyyy4977 thought the picture looked fake. Source: @jovial_dolphin_29796, 2026-07-28
  Source: @sssyyy4977, 2026-07-28
• @huzifei talked to the Xianyu seller, who said he runs it at 72 GB and cannot guarantee stability, and offered to teach the method for 50 RMB. @v1ndim claimed the seller unlocked his card to 80 GB and said to "follow Jon Pry's approach", but no model run above 64 GB or clean gpu-burn result was posted. @adih1266 linked zerobitcopy/cmp-easyunlock as a possible source (above). Source: @huzifei, 2026-07-28
  Source: @adih1266, 2026-07-28
  Source: @v1ndim, 2026-07-28
• @_studebaker_: assume nothing new until proven, especially from someone selling overpriced 10 GB cards into the hype; @farawaythrower called it a scam. Source: @_studebaker_, 2026-07-28
  Source: @farawaythrower, 2026-07-28

Contested 10 GB

"80 GB unlock" videos and fake-capacity resale

Warning 10 GB 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 linked his earlier message on 10 GB instability (Discord), Capone said addressing more than 40 GB crashes, Studebaker suggested asking the creator to write past 40 GB, and DeeJay noted the video only used about 27 GB. barackorama warned that a card can be made to report 80 GB even when it is not usable, so scammers can sell it the way fake-capacity SSDs are sold; 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 data above 40 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 listing: "80g unlock with 60-70g usable" (September 2026)

On 2026-09-18 Arcanum found a Xianyu (Goofish) listing that used a screenshot involving a member of this server and claimed an 80 GB unlock for the 10 GB card with 60 to 70 GB usable; Arcanum was unsure what that meant. insomniac pointed out a red square in the screenshot that suggested it had been edited. Members noted that Chinese sellers watch this server. An unverified seller claim.

Reported 10 GB
Source: @_arcanum_, 2026-09-18
Source: @ins_mniac, 2026-09-18
Source: @giaki3003, 2026-09-18

Seller claims of a 60 GB unlock (September 2026)

Warning 10 GB 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 said no public unlock goes beyond 40 GB on these cards, and Arcanum questioned how it could work physically. Treat seller claims of a 60 GB unlock as unverified when pricing 10 GB cards.

Source: @tokukarin, 2026-09-22
Source: @lycorascence, 2026-09-22
Source: @_arcanum_, 2026-09-22

128 GB and 140-SM screenshots are AI mockups or edits

Warning 10 GB On 2026-09-05 giaki3003 posted an nvidia-smi image of a 10 GB card showing 128 GB and 140 SMs and said it was a mockup he had asked an AI to make; MaitOPS complained that it had been pasted in many channels whenever people discussed unlocking more HBM. regaly explained why 128 GB cannot exist (GA100 has at most six stacks of at most 16 GB each) and posted a fake 256 GB screenshot to show how easily such output is edited; he had also seen "128 SM unlocked" CMP 170HX screenshots on Alibaba, which he thought probably fake. A screenshot of nvidia-smi is not evidence of an unlock.

Source: @giaki3003, 2026-09-05
Source: @reggie7756, 2026-09-05
Source: @_maitops, 2026-09-05
Source: @_maitops, 2026-09-05

Open leads

Untested: read the MMU lock registers on a 10 GB card

In reply to the 80 GB DMA gist on 2026-09-30, okrasit. asked whether anyone had read the NV_PFB_PRI_MMU_LOCK address registers (low and high) and the lock's privilege-level mask on 10 GB cards. The implication is that an MMU-locked range might explain the DMA ceiling at about 40 GiB. Nobody answered in this history.

Theory 10 GB
Source: @okrasit., 2026-09-30

Lead: symbolise the recurring GSP fault addresses

cuddylac997, posting his AI agent's analysis on 2026-09-30, proposed checking whether the 40 GB frame-buffer cap comes from a separate CMP GSP firmware binary (then stock GA100 firmware might help) or from the same firmware branching on device ID 0x2082 or on fuses. The key step: symbolise the recurring GSP fault addresses 0x5b2ba28 and 0x5335ab0 against the exact GSP build, to tell whether the cap is gated by device ID or fuses (then a spoof or an A100 route would be viable) or is hard. He asked where A100 firmware had been posted; nobody answered in this period. The ggualerz reports above list device-ID overrides among their dead ends.

Theory 10 GB
Source: @cuddylac997, 2026-09-30

Request: the evidence behind the 2026-07-14 FBPA override attempt

On 2026-10-02 cuddylac997 asked who knows about the Consensus-Protocol/cmp170hx write-up and requested the original evidence from an attempt dated 2026-07-14: the patch or command, the register dumps before and after, and the device, boot and driver identity. The write-up says a software override of the FBPA configuration left the effective status unchanged, but it is secondary synthesis, and another of its pages contradicts its test history. He also asked about another attempt from June, and said explicitly not to repeat the write from the summary; segmond said he knows about it. As a timeline marker: a software FBPA override was reportedly tried around 2026-07-14 without changing the status.

Contested 10 GB
Source: @cuddylac997, 2026-10-02
Source: @cuddylac997, 2026-10-02
Source: @segmond, 2026-10-02

Related

10 GB card, Memory unlock on the 10 GB card, Which card do I have, Variant comparison, Verifying an unlock, Testing and burn-in, Xid 31 and the reserved top of VRAM, RIR fuse burn, Full SM unlock, Fuses and floorsweep maps, Gpcprobe, Driver unlock, Cmpunlocker forks, Unlock history: who found what, and when, Odd variants and engineering samples, 32 GB cards (subsystem 10de:1586), Price history, July to October 2026, Price history and market, Buying from China as a foreigner, Rumours and claims to verify, Reference library.

Imported from the cmpunlocker Discord with credit; see Credits.