10 GB card

Reported 10 GB

The 10 GB card is the CMP 170HX with PCI ID 10de:2082 and Samsung HBM2e. It has less memory than the 8 GB card after unlock, but its RIR fuse records are unused, which opens a path to most of a full GA100: up to 128 SMs, PCIe Gen3 and more. Plan on 40 GB of memory; 48 GB is a lottery. Not sure which card you have? See Which card do I have. New owner? Start with Start here#10gb.

At a glance

	Stock	cmpunlocker only	After the RIR burn and driver work
Memory	10 GB	40 GB (40960 MiB; about 37 to 38 GiB usable)	40 GB; 48 GB if the sixth stack is healthy; 80 GB not achieved
Memory type	Samsung HBM, five active stacks on a 5120-bit bus, about 1215 MHz
SMs	70	74	up to 128, per card (112, 122 and 128 seen)
PCIe	Gen1 x4	Gen2 x4	Gen3 (validated); Gen4 reported, not public; x16 needs the PCIe capacitor mod
ECC	off	on	on
NVLink	fused off	none	not public; needs a board-level mod
Windows	no	via the A100 40G / A800 40G driver	same

Unlock geometry per the cmpunlocker constants. Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub, posted by @rh3dai, 2026-10-08 The 10 GB card has 16 unblown RIR rows, which allow a full GA100 configuration while stable memory stays at 40 or 48 GB. Source: @wakeupmbrc1, 2026-10-07 Unlocked 10 GB cards land at 112, 122 or 128 SMs. Source: @wakeupmbrc1, 2026-10-27 A 10 GB card now equals an A100 40G PCIe with more SMs. Source: @ob88594, 2026-10-23 Each number carries its status and sources in the sections below: 40 GB and the post-burn SM and Gen3 results are reproduced by several owners, 48 GB is one-card-in-five, and 80 GB is not achieved: cards that report 80 GB fold or crash above about 40 GB (see #Memory and The 10 GB card: provenance and the 80 GB question).

Proven Memory vendor and bus. Every 10 GB card carries Samsung HBM and every 8 GB card SK Hynix: Mr Pec stated it and RazK confirmed it on 2026-07-19, and Luna, supasnipa360 and Arcanum repeated it the next day. Thanos counted four working stacks on the 8 GB card and five on the 10 GB card, which gives coolgod's 4096-bit and 5120-bit buses. Arcanum: the 10 GB card's HBM runs at about 1215 MHz against about 1458 MHz on the 8 GB card, so the fifth stack roughly evens out the bandwidth. Whether the Samsung stacks are HBM2 or HBM2e is a separate debate (Which card do I have#Memory vendor). Source: @mrpec, 2026-07-19
Source: @port00, 2026-07-19
Source: @thanos8912, 2026-07-19
Source: @coolgod_72592, 2026-07-19
Source: @lunaroseoc, 2026-07-20
Source: @supasnipa360, 2026-07-20
Source: @_arcanum_, 2026-07-20

Reported What is public and what is still gated. Miséricorde's case for the 10 GB card: a fully unlockable GA100 with up to 128 SMs, PCIe 4.0 and NVLink, and 48 GB on some cards against 40 GB by default; 80 or 96 GB is not possible without Samsung internals. In the same thread mrpandagod ran 122 SMs with PCIe 3 patches and eXile's 10 GB card ran Gen2 x16 after the capacitor mod. The 48 GB, NVLink and PCIe 4.0 pieces are not in the public cmpunlocker release. Humvee55 first announced the extra SMs, the sixth stack and Gen3/4 on 2026-09-29 (see #Step 2: the RIR burn). Source: @wakeupmbrc1, 2026-10-12
Source: @djpandaman, 2026-10-12
Source: @exile_1ck, 2026-10-12

Unlock results

Details and steps: Driver unlock and Unlocks.

Step 1: cmpunlocker (same as the 8 GB card)

With the cmpunlocker repo a 10 GB card gets 40 GB and a few extra SMs; unlocking the whole card takes real legwork. Source: @lycorascence, 2026-10-16 v0.5 gives +4 SMs and ECC on both variants. Source: @lia_oc, 2026-10-22 Driver 615.71.09 works with v0.5 on 10 GB cards. Source: @68x0, 2026-10-05

• Proven 40 GB from the first release. On 2026-07-19 Far reported his 10 GB cards (10de:2082) working properly at 40 GB, with no PCIe Gen2 yet, 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, 80 GB on the 10 GB card possible but unstable. meatsus and giaki3003 noted that the VRAM refresh-rate change from the original paper can be applied, but the result is not stable enough. 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: @giaki3003, 2026-07-21
  Source: @meatsus, 2026-07-21
• Reported "40 usable, 80 detectable". 40 GB stayed the working figure from July to early September 2026. Eksitus's card reported 40960 MiB in nvidia-smi on 2026-07-22, with a 250 W power limit and 42 W in P0 at 38 C idle. DaLZy had seen no 10 GB card that failed to unlock to 40 GB (2026-07-27). CUTE's summary on 2026-07-31: 10 GB cards unlock to 40 GB, and 80 GB seemed possible but did not work yet. RedPandaMining's six 10 GB cards ran the 40 GB unlock on 2026-08-26. In September Far answered a newcomer's question with "40 usable, 80 detectable". Source: @eksitus, 2026-07-22
  Source: @dalzyyy, 2026-07-27
  Source: @cute_sc2, 2026-07-31
  Source: @redpandamining, 2026-08-26
  Source: @farawaythrower, 2026-09-11
• Proven How much of the 40 GB you can use. nvidia-smi shows 40960 MiB, but less of it is usable. On 2026-08-06 arabus reported both of his 10 GB cards unlocked "perfect" and running as 40 GB cards, and MrManny confirmed 10 GB to 40 GB as the stable unlock. cuddylac997 (2026-08-13): a model that goes over the cap of about 37 GB crashes one card and needs a restart; cmpunlocker2 avoids the restart. His keeper leaderboard for four 10 GB cards showed 33.0 GiB verified clean per card (2026-08-23), and his four-card server gives about 37.7 GiB usable per card (160 GB pooled). Lycorascence (2026-08-20): exhausting all VRAM stalls his 40 GB card until the whole machine is rebooted; see Xid 31 and the reserved top of VRAM. Plan on about 37 to 38 GiB per card after reservations. Source: @arabus_74064, 2026-08-06
  Source: @mrmannyatx, 2026-08-06
  Source: @cuddylac997, 2026-08-11
  Source: @cuddylac997, 2026-08-13
  Source: @lycorascence, 2026-08-20
• Reported HiveOS shows the cards as "malfunction". With the early unlocker on 2026-07-18, GMZombie007 got only the first card working and HiveOS marked the cards "malfunction"; Amogh confirmed the unlocker had no multi-card support yet. After running remove.sh for a fresh install and rebooting, both 10 GB cards came up working at 40 GB. Aluminer saw the same status on 2026-07-28; SealClubber: it only means a bad driver, the cards are fine. Aluminer reflashed HiveOS, started over, and the memory unlock applied. Source: @gmzombie, 2026-07-18
  Source: @_j3ns, 2026-07-19
  Source: @aluminer_, 2026-07-28
  Source: @gmzombie, 2026-07-28
  Source: @sealclubber, 2026-07-28
• Reported Gen2 fails when 8 GB cards share the box. orrorin's 10 GB card would not unlock to Gen2 while 8 GB cards were also installed, and at one point it sat at PCIe 1.0; unplugging the 8 GB cards and re-running the unlock fixed it. On 2026-09-17 he narrowed it down: Gen2 link training fails at boot, but a hard reload of the driver after boot makes it work. See Gen2 link retrain. Source: @_orrorin, 2026-09-16

Step 2: the RIR burn

Warning This is permanent. It burns two of the card's 16 RIR fuse records and cannot be undone. Read RIR fuse burn first.

• What it does: two RIR records, 0x0015 and 0x0095, clear the DISABLE_SW_OVERRIDE fuse (170hx-rir-tool writes them to RIR row 0 as the word 0x00950015); after that, SMs, PCIe speed and the sixth stack can be overridden in software. lxwx confirmed the record pair on a card on 2026-10-11. Source: @lxtend, 2026-10-11 In mrpandagod's words: a 2-record burn that allows the SM unlock, PCIe 3.0/4.0 and the sixth stack on some cards. Source: @djpandaman, 2026-10-12
• Tool: 170hx-rir-tool (lxwx, MIT) burns the records and dumps the fuses; it needs iomem=relaxed and root. Source: repo github.com/lixiangwuxian/170hx-rir-tool by lixiangwuxian on GitHub, posted by @lxtend, 2026-10-10 It also needs the cmpunlocker-patched driver (or a 610+ driver) loaded. Source: @lycorascence, 2026-10-20 To unbind just the CMP without taking down other GPUs, write to that GPU's sysfs unbind path instead of unloading the driver. Source: @c133, 2026-10-18
• Proven Burned and working for several owners: after lxwx's card, GhostInTheDatacenter on 2026-10-19 ("Finally!!!" with a screenshot), Source: @ghostinthedatacenter_87281, 2026-10-19 orrorin on 2026-10-20 ("we did it boys" with a screenshot; the text names no link speed or SM count). Source: @_orrorin, 2026-10-20
• Warning Getting the packing wrong wastes records or bricks the card. Targeting the wrong address has a good chance of bricking. Source: @djpandaman, 2026-10-05 An earlier attempt with a different encoding (0x15011401) changed nothing and wasted a row. Source: @lxtend, 2026-10-10
• Do you need to burn at all? Contested In the order posted:
  • mrpandagod: going from 74 to 80 SMs needed no fuse override; 80 to 122 definitely did. Source: @djpandaman, 2026-10-10
  • OB88594: the 10G card only needs the software overlay and no writes to other RIRs. Source: @ob88594, 2026-10-26
  • Humvee55: a recipe plus software IFF was shared in #general; use that and do not waste RIR slots. Source: @scallion_9883, 2026-10-26 Earlier he posted his RIR and software-IFF notes (a 10 KB text file) for an owner who had burned the override fuse but could not finish the unlock; Feralzi fed them to his agent and reached 126 SMs. Source: @scallion_9883, 2026-10-16
    Source: @feralzi, 2026-10-17
    Source: @feralzi, 2026-10-17
• Proven Why only the 10 GB card can take the burn. On 2026-09-29 CUTE stated that all 16 RIR records on the 10 GB card are unused, so one can be spent to unlock the card, while on the 8 GB card 15 are burnt and the last one is "poisoned"; his explanation is that NVIDIA forgot to burn the RIRs on the 10 GB cards in manufacturing. Humvee55: nowhere else in Ampere is there a single instance of fully available RIRs, and the other A100 variants he checked have all RIRs trashed. Amogh confirmed it is real and stressed that an RIR record should be spent only to disable the software-override fuse, because records are one-time and limited; everything else is then changed through the resettable fuses. N2DCRYPT expanded RIR as "Redundant Information Row". Source: @_j3ns, 2026-09-29
  Source: @cute_sc2, 2026-09-29
  Source: @scallion_9883, 2026-09-29
  Source: @n2dcrypt, 2026-09-29
• Reported RIR read-out on both variants. Zarquon read the RIR records on 2026-09-28: on a 10 GB card all 16 records are blank, with nothing programmed; on an 8 GB card the record rows are invalidated except for one active factory record that sets a poison-disable fuse. Big Ptoughneigh warned that any write test must be safe for both the RIR and the main fuse array or it can brick the card. Source: @zarquon3157, 2026-09-28
  Source: @bigptoughneigh, 2026-09-29
• Proven The real split between the variants is RIR occupancy. The software-override gate is the same on both variants: the override-disable fuse is set on every card of a 15-card cohort that included both 170HX variants, with the same PCIe fuse fingerprint and no PCB difference. What differs is the RIR macro: the Hynix 8 GB card (10de:20c2) ships with it consumed (15 of 16 slots invalidated, one live record, none free), the Samsung 10 GB card (10de:2082) ships blank with 16 free slots. The burn tool refuses non-blank slots, so the gate-clear record has nowhere to go on an 8 GB card. Mark Linton's read-only dumps of two 8 GB cards (2026-10-10, no burns) and GitHub reports on further 8 GB cards show the same picture. Humvee55 was the first to say NVIDIA left the 10 GB card's RIRs empty (next item). Source: @marklintonca, 2026-10-10
  Source: repo github.com/ahnguyen17/ga100-fuse-re by ahnguyen17 on GitHub
  Source: repo github.com/lixiangwuxian/170hx-rir-tool by lixiangwuxian on GitHub
  Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub
• Reported Humvee55's announcement, 2026-09-29. 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, the extra memory stack (up to 48 GB if that stack is not marked defective) and PCIe Gen3/4 all work: for compute the 10 GB card wins, for memory the 8 GB card (64 GB) stays 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. Source: @scallion_9883, 2026-09-08
  Source: @scallion_9883, 2026-09-29
  Source: @reala7096, 2026-09-29

SMs

Details: Full SM unlock.

• mrpandagod took one card from 74 to 80, 94, 108 and then 122 SMs. Source: @djpandaman, 2026-10-09 lxwx got 6 more by overriding OPT_TPC_DISABLE, Source: @lxtend, 2026-10-12 then all 128 SMs at full clock on 2026-10-17. Source: @lxtend, 2026-10-17 Feralzi reached 126. Source: @feralzi, 2026-10-17 Waypoint enabled every SM not marked defective. Source: @waypoint3642, 2026-10-17 Humvee55 runs 20 cards at 128 SMs and Gen3 x16. Source: @scallion_9883, 2026-10-16
• Defective GPCs. GPCs marked defective really are defective. Source: @zorg33, 2026-10-22 You can still enable a defective GPC if you mask off the defective TPCs inside it. Source: @djpandaman, 2026-10-22 Colo let an agent A/B test SM pairs and disabled only the faulty ones; one card runs 126 tested SMs. Source: @colo2442, 2026-10-22
• How far a given card goes. Contested Details on Full SM unlock#How many SMs a 10 GB card reaches. In the order posted:
  • xeax: a 10 GB card can enable every SM that is not defective, up to 128 if you are very lucky; the work is per card: unlock it with the RIR tool, test which locked parts work, then build a custom mask for that card. Source: @xeax, 2026-10-25
    Source: @xeax, 2026-10-25
  • orrorin: both of his 10 GB cards stop at 80 SMs; anything above 80 failed a real model load, also with the other locked GPC. Source: @_orrorin, 2026-10-25
    Source: @_orrorin, 2026-10-25
  • OB88594: more than 100 SMs is very common, so this should be an operating error; xeax: more likely heavily binned dies. Source: @ob88594, 2026-10-25
    Source: @xeax, 2026-10-25
• Reported Xid 109 at 128 SMs. With defective GPCs enabled, a workload crashed with Xid 109 CTX SWITCH TIMEOUT; Source: @lxtend, 2026-10-17 the fix was to disable the enabled-defective GPCs one at a time until stable; Source: @wakeupmbrc1, 2026-10-17 disabling GPC1 fixed it. Source: @lxtend, 2026-10-17
• Reported Silicon lottery: defective versus disabled. Whether a card's inactive GPCs and stacks are marked defective or only disabled decides how far it goes. Waypoint's 10 GB card has one defective GPC, so it cannot reach 126 SMs; Lycorascence's card has its inactive GPCs and HBM stacks only disabled; GhostInTheDatacenter judged zero-defect cards to be rare. matstatman warned that not every card goes that high: "you might get 40gb or a brick". OB88594: 48 GB is probabilistic and only some cards reach it. From a small sample, mrpandagod reported on 2026-10-01 that 10 GB cards typically have 2 of 12 FBPs fused off and marked defective, while an 8 GB card he checked had 2 defective and 2 disabled; Lycorascence noted those two FBPs are marked disabled instead of defective on some 10 GB cards. Check GPC and stack status with Gpcprobe before paying a 10 GB premium. Source: @waypoint3642, 2026-09-29
  Source: @lycorascence, 2026-09-29
  Source: @ghostinthedatacenter_87281, 2026-09-29
  Source: @matstatman, 2026-09-29
  Source: @ob88594, 2026-09-29
  Source: @djpandaman, 2026-10-01
  Source: @lycorascence, 2026-10-01

PCIe

Details: PCIe Gen3 on the 10 GB card.

• Gen3. lxwx had Gen3 working on 2026-10-12 (his slot is x4, so Gen3 x4 equals Gen1 x16 bandwidth). Source: @lxtend, 2026-10-12 A wrapper tool validated Gen3 at boot on two 40 GB cards with a 30-minute stress test and no errors. Source: repo github.com/evanrodgers/cmp-170hx-rir-burn-util by evanrodgers on GitHub, posted by @hezz1891, 2026-10-26
• Gen3 how-to (summary of lxwx's guide). For the 10 GB card (10de:2082) on open kernel modules 615.71.09. Steps, in short:
  1. Burn the two RIR records with 170hx-rir-tool so the software-override gate opens (see Step 2 above).
  2. Build the driver with the guide's patch: a probe-time fuse override that clears the two Gen3-disable fuses, plus a Gen3 register sequence run after GSP boot.
  3. Load it with options nvidia NVreg_RegistryDwords="RmForceEnableGen2=1;RMPcieLinkSpeed=0x5;CmpGen3=1" cmp_fuse_gen3=0.
  4. Train the link one step at a time, Gen1 to Gen2 to Gen3, with the guide's speed tool and script.
  5. Accept it only when lspci -vv shows LnkSta: Speed 8GT/s on both ends and a load test gives no Xid or AER errors.

  cmp-pcie-speed 2 drops back to Gen2 at runtime. One Gen3 transition took a card off the bus until a power cycle. Only 10de:2082 was validated. Full guide: Gen3_unlock_en.md. Source: repo github.com/lixiangwuxian/170hx_tips by lixiangwuxian on GitHub, posted by @lxtend, 2026-10-19

• Gen4. mrpandagod has Gen3 and validated Gen4 behind a PLX switch. Source: @djpandaman, 2026-10-09 The 10 GB card can actually use PCIe 4. Source: @cute_sc2, 2026-10-15 No public Gen4 recipe exists yet. Without the capacitor mod his Gen4 attempts kept falling back to 8.0 GT/s. Source: @djpandaman, 2026-10-08
• Width. x16 at any generation needs the PCIe capacitor mod. Source: @lxtend, 2026-10-17
• Reported Driver crashes from the power feed, not the link. Lycorascence's 10 GB card kept crashing the driver and blocking his other GPUs while it was powered from a secondary CPU (EPS) line on his PSU. Moving it to a PCIe Y-splitter cable and a different riser board fixed it: the card stayed up overnight and passed a 5-minute Vulkan memtest at 40 GB (2026-08-13, reconfirmed 2026-08-16). Which of the two changes mattered was not isolated. See Power supplies and cabling and Risers and bifurcation. Source: @lycorascence, 2026-08-13

Windows

The unlock works on Windows on the 10 GB card with the A100 40G or A800 40G driver. Source: @wakeupmbrc1, 2026-10-11 A screenshot of a working 10 GB unlock on Windows was posted on 2026-10-16. Source: @arabus_74064, 2026-10-16

Memory

Details: Memory unlock on the 10 GB card.

• Layout. 20 of 24 FBPA active (5120-bit) with 1 GiB dies; ceiling 48 GiB. Source: repo github.com/matstatman/gpcprobe by matstatman on GitHub, posted by @matstatman, 2026-10-20 Every one of six cards bigfoot tested had exactly two defective FBPs. Source: @bigfoot1144, 2026-10-09
• Proven 40 GB is where the unlock lands and where most cards stop. Source: @zorg33, 2026-10-07
  Source: @wakeupmbrc1, 2026-10-07
  Source: @programmerboi, 2026-10-25
• Reported 48 GB needs the sixth FBP to be good, Source: @wakeupmbrc1, 2026-10-17 about 1 card in 5: mrpandagod put it near 20%, Source: @djpandaman, 2026-10-12 6 of 30 cards he saw tested could do it, Source: @djpandaman, 2026-10-17 and gpulover gives 20% or less. Source: @artistosu, 2026-10-26 Check yours with gpcprobe before you count on 48 GB (Which card do I have#Check the memory layout).
• How often is the sixth stack usable? Contested In the order posted:
  • bigfoot: all six of his 10 GB cards have the last stack marked defective. Source: @bigfoot1144, 2026-10-19
  • Lycorascence: the last stack on his card is marked disabled, not defective. Source: @lycorascence, 2026-10-19 He also described the card as having "16/16 fuses" with the GPCs only disabled; that count came from gpcprobe's RIR checker, since he could not dump the RIR itself, and mrpandagod questioned it on those grounds. Source: @lycorascence, 2026-10-20
    Source: @lycorascence, 2026-10-20
    Source: @djpandaman, 2026-10-20
    Source: @lycorascence, 2026-10-20
  • lxwx: defective-labelled stacks are mostly good anyway (untested). Source: @lxtend, 2026-10-20
• 80 or 96 GB. Contested Not achieved; cards that read 80 GB were unstable. Source: @tekkenz, 2026-10-12 In the order posted:
  • The extra capacity is aliasing (folding) unless address bit 14 is freed by a further unlock. Source: @colo2442, 2026-10-14
  • The apparent capacity is aliasing in L2. Source: @scallion_9883, 2026-10-16
  • The stacks are binned 16 GB HBM2e and the upper half would need Samsung vendor codes. Source: @djpandaman, 2026-10-17
  • Full summary: Which card do I have#80 GB on the 10 GB card and Memory unlock on the 10 GB card.
• Memory vendor. Reported Samsung HBM2e (the 8 GB card uses SK hynix). Source: @darkmaniac7, 2026-10-30 (lab notes) Earlier posts disagreed on HBM2 versus HBM2e for these Samsung stacks; they are listed under the fact on Which card do I have#Memory vendor.
• Bandwidth. Theory With all six stacks the bus is 6144-bit, which could pass 2.5 TB/s, or 3 TB/s if Samsung memory overclocks like Hynix; untested. Source: @wakeupmbrc1, 2026-10-07 Measured: an unlocked 10 GB card reached 1,741 GB/s at full stock. Source: @cuddylac997, 2026-10-07 clpeak on mrpandagod's 122-SM 40 GB card measured 1.68 to 1.82 TB/s global memory bandwidth; full results on Benchmarks#Compute and memory peaks. Source: @djpandaman, 2026-10-23
• Temperature. The 10 GB memory runs a few degrees warmer than the 8 GB at the same idle power (41 C versus 37 C after a repaste, both near 33 W). Source: @ghostinthedatacenter_87281, 2026-10-13 It has two more half stacks active, so that is expected. Source: @colo2442, 2026-10-13

History, July to October 2026

The short version of what was tried on the 10 GB card's memory before the RIR burn. The full record, with every thread, is on The 10 GB card: provenance and the 80 GB question. For the 40 GB baseline in this period see #Step 1: cmpunlocker (same as the 8 GB card).

July 2026

• Theory Bin cards by VID. Far suggested reading each card's VID (voltage ID) to find the cards that would hold 80 GB, since silicon quality varies card to card. Untested. Source: @farawaythrower, 2026-07-19
• Slower memory or faster refresh. Contested In the order posted:
  • Samuel Hayden: 80 GB could be reached by downclocking the memory. Source: @samuel_hayden., 2026-07-24
  • btakk_ak47: unverified, and no driver-level way to lower the memory frequency had been found. Source: @btakk_ak47, 2026-07-24
  • okrasit. quoted a September research document: the full 80 GB is physically present but shows 2,796 errors under stress at the stock refresh interval and 0 errors when refreshing more often, at about 32% throughput cost. CUTE: lowering the HBM speed did not fix the folding. Source: @okrasit., 2026-09-29
    Source: @cute_sc2, 2026-09-29
• Would the missing VRM parts stabilize 80 GB? Contested In the order posted:
  • tugoubxs and Matty Ice hoped that fitting the unpopulated inductors, DrMOS and capacitors would supply more power and stabilize 80 GB. Source: @tugoubxs, 2026-07-24
    Source: @icematty, 2026-07-24
  • meatsus: the GPU does not sense the VRM phase count and the existing VRM delivers the roughly 250 W full-load draw; he and SealClubber put the instability down to silicon that missed the A100 standard. The thread's consensus was no. Source: @meatsus, 2026-07-24
    Source: @sealclubber, 2026-07-24
• Theory Open questions from July. ti asked whether the 80 GB results in the original paper could be reproduced at all if the card freezes as soon as VRAM use passes 40 GB; W.H. believed a working method exists but was never released, and tugoubxs noted that nobody had claimed a stable 80 GB. W.H. also observed that the 10 GB card's VRAM addressing appears to start at 1G rather than at 0, and gk_hun theorized 10 GB sub-variants with differently packed HBM. All untested. Source: @jovial_dolphin_29796, 2026-07-24
  Source: @winghorizon., 2026-07-24
  Source: @tugoubxs, 2026-07-24
  Source: @gk_hun, 2026-07-24
• Reported No quick path to 80 GB. Samuel Hayden (2026-07-26): the HBM control registers could not be reached with the exploit as it stood, so there was likely no quick path to 80 GB (or to ECC); anyone buying a 10 GB card should not expect more than 40 GB any time soon. Source: @samuel_hayden., 2026-07-26
• Warning "80 GB unlock" videos and fake-capacity resale. A YouTube video titled "I Unlocked 80GB VRAM on a Mining Card!" used only about 27 GB (DeeJay, 2026-07-29), and Capone said addressing more than 40 GB crashes. barackorama warned that a card can be made to report 80 GB when that memory is not usable, so scammers can sell it the way fake-capacity SSDs are sold. Test any card sold as "80 GB" by writing and verifying data above 40 GB. See Rumours and claims to verify. Source: @deejay9983, 2026-07-29
  Source: @beefbonecapone, 2026-07-29
  Source: @barackorama, 2026-07-29

August 2026

• Theory Binning argument. Jeremy (Ends): a die with enough working HBM for a stable 80 GB would have been sold as an A100 80GB, so stable 80 GB cards are statistically unlikely. prometasfw countered that a die can have good stacks but failed SMs, which could not become an A100, and Humvee55 has a 10 GB card with only 56 active SMs. Source: @endsjeremy, 2026-08-06
  Source: @prometasfw, 2026-08-06
  Source: @scallion_9883, 2026-08-06
• Reported 80 GB in another server. Rubbra_F said on 2026-08-09 that someone got 80 GB on the 10 GB card in the separate "clean room" Discord server; ju5t5omedude said similar work had been done there about a month earlier. Those posts cannot be checked from this server's history. Source: @rubbra_f, 2026-08-09
  Source: @ju5t5omedude, 2026-08-09
• Is the upper 40 GB there at all? Contested In the order posted:
  • giaki3003: the 10 GB cards people are getting have only 40 GB physically; an 80 GB-capable card may exist, but he had never seen one. Source: @giaki3003, 2026-08-10
  • Massimo and Xeax: the card has 80 GB physically with only 40 GB stable; Xeax adds that one person claims stability above 40 GB by giving up about 30% memory throughput, but most cards are binned, so results vary. Source: @mv.nyc, 2026-08-10
    Source: @xeax, 2026-08-10
  • cuddylac997: a firmware limit, not binning, because every card stops at exactly 40 GB. Source: @cuddylac997, 2026-08-10
  • irythros: 80 GB can be reached but the memory is corrupted, citing a report of addresses wrapping after 40 GB and silently overwriting earlier data. Source: @irythros, 2026-08-21
  • xanthe_: the original research paper describes 80 GB first unstable and then made stable by a tweak at some performance cost. Source: @xanthefin, 2026-08-21
• Theory Picking cards by serial number. Studebaker suggested buyers may be selecting the better 10 GB cards by serial number, as happened with the CMP 100-210. No evidence was posted and no serial ranges are known to map to better cards. Source: @_studebaker_, 2026-08-10
• Can 80 GB work at all? Contested In the order posted:
  • Luna: 80 GB on the 10 GB card works, but only about two fully functional cases had been recorded and the top one or two stacks are usually defective. Source: @lunaroseoc, 2026-08-16
  • LightOverDarkness: without NVIDIA's signing key or fuse authority, linear 80 GB is not software-unlockable. Source: @light_2212, 2026-08-16
• Samsung cards and Xid 79. Contested The ahnguyen17/cmp-170hx-vllm README (https://github.com/ahnguyen17/cmp-170hx-vllm) claims Samsung-memory cards needed exllama instead of Marlin and eager mode and still hit Xid 79 under sustained load. In the order posted:
  • Petri Krohn doubted it, since both variants share GA100 silicon; giaki3003 dismissed the README. Source: @petrikrohn_26497, 2026-08-25
    Source: @giaki3003, 2026-08-25
  • m3gagluk: the real difference is 40 GB against 64 GB configurations on a driver without the memory reservation fix; Petri agreed that 40 GB cards can hit the reservation at the top of memory. Source: @m3gagluk, 2026-08-25
    Source: @petrikrohn_26497, 2026-08-25

September and early October 2026

• Reported What the fuses do to the HBM. Petri Krohn: the 10 GB card has two FBPs (half-stacks) marked defective or intentionally disabled by fuses. The fuses do not physically separate the HBM from the GPU die; they only set a register that tells the firmware which parts of the package to ignore. Source: @petrikrohn_26497, 2026-09-17
• 48 GB at most? Contested In the order posted:
  • Humvee55: no 10 GB card runs stable at 80 GB; at most one extra stack can be unlocked, for 48 GB, on some "unicorn" cards. Source: @scallion_9883, 2026-09-18
  • gpcprobe on orrorin's and Lycorascence's cards read six stacks (four full, two half), 20 of 24 FBPA, 40 GiB exposed and 48 GiB at most. Source: @matstatman, 2026-09-18
    Source: @_orrorin, 2026-09-18
    Source: @lycorascence, 2026-09-18
  • MaitOPS: 80 GB of HBM is inside and readable and writable from the SMs, but normal workloads crash the GSP firmware. Source: @_maitops, 2026-09-18
  • CUTE: 10 GB cards "as of now" carry no more than 48 GB of HBM on board, so the maximum is six 8 GB stacks and only if no HBM die is defective; Con3po 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
• Reported A Xianyu listing claiming 80 GB with 60 to 70 GB usable. Arcanum found it on 2026-09-18; insomniac pointed to a red square in its screenshot that suggested an edit. Unverified seller claim. Source: @_arcanum_, 2026-09-18
  Source: @ins_mniac, 2026-09-18
• Reported Where the crash comes from. MaitOPS (2026-09-19): Jon Pry was the first to find the 80 GB of HBM on the 10 GB card; the crash hits the same point on multiple cards, and his own AI-assisted reverse engineering points to a bug in the GSP memory allocation map, which could not be patched at the time. Source: @_maitops, 2026-09-19
• Reported Layout readout. Tokukarin's reading of a 10 GB card: locked, 20 of 24 FBPA (5120-bit), 10 GiB exposed and 12 GiB at most; unlocked to 40 GB, 1 GiB per die, 40 GiB exposed and 48 GiB at most. Lycorascence: no public unlock goes beyond 40 GB, and nobody knows how many 10 GB cards are truly 80 or 96 GB capable. cuddylac997's cards show the same 20-active-FBPA baseline. Source: @tokukarin, 2026-09-21
  Source: @lycorascence, 2026-09-21
  Source: @cuddylac997, 2026-09-30
• cuddylac997's hbm45 test. Contested In the order posted:
  • cuddylac997: with the 80 GB geometry patch, a CUDA kernel test that writes a unique pattern to every chunk and verifies after all writes came back clean to at least 76 GiB on three 10 GB cards, with no folding. Host-to-device copies above about 40 GiB fault in GSP DMA, so in his reading 80 GB is physically real but needs a GSP DMA fix or loaders that write weights from kernels. His gist: https://gist.github.com/cuddylac997/c3d80faa2430e3650cd934eda5fd65a9. When he first joined, two months earlier, he had got 76 GB to show up and was talked out of continuing. Source: @cuddylac997, 2026-09-30
    Source: @cuddylac997, 2026-09-30
  • Sojun: the memory is accessible but not stable; CUTE and mrpandagod had reported folding (see the aliasing positions under 80 or 96 GB above). Source: @sojun80, 2026-09-30
• Reported "Very unlikely". GHOSTi posted his card's fuse-probe output and asked on 2026-10-02 whether it would reach 80 GB; Colo answered "very unlikely", and CUTE said 80 GB is possible only if the memory is physically there. Source: @ghost9191, 2026-10-01
  Source: @colo2442, 2026-10-01
  Source: @cute_sc2, 2026-10-01
• Theory The July 14 attempt. On 2026-10-02 cuddylac997 asked for the original evidence from the July 14 Consensus Protocol attempt at 80 GB, to understand why it failed; segmond said he knows about it. Nothing more was posted. Source: @cuddylac997, 2026-10-02
  Source: @segmond, 2026-10-02

Limits

• Power. Reported Above 80 SMs the card hits its power cap and memory clocks drop. mrpandagod's 122 SM card throttled even with the core locked at 1,095 MHz. Source: @djpandaman, 2026-10-12 The cap triggers at 82 SMs, the moment an SM on a new GPC is enabled, and memory is forced to 1,100 MHz. Source: @djpandaman, 2026-10-13 Self-reported power becomes unreliable at that point (400 W+ readings on a 250 W VBIOS). Source: @colo2442, 2026-10-22 Contested Where it starts: xeax's (untested) sweep README treats power readings that jump above about 90 SMs as a telemetry artifact; Source: @xeax, 2026-10-22 mrpandagod answered the same day that the wattage issues start after 80 SMs, not 90, "just when you enable an SM on a disabled GPC". Source: @djpandaman, 2026-10-22 Warning On 2026-10-23 mrpandagod's 122-SM card spiked briefly to about 400 W during benchmarks with the power limit at 250 W; he was not sure whether the value was measured or calculated. xeax, who is weighing whether to fit the 6 MOSFETs missing from his card, fears overloading the VRM. Safe limits for fully SM-unlocked cards are not known; see Power supplies and cabling#PSU sizing. Source: @xeax, 2026-10-23
  Source: @djpandaman, 2026-10-23
• Raising the limit. Contested Approaches differ, in the order posted:
  • Give the driver a truncated 8 GB unlocked VBIOS; this reached 300 W. Source: @bigptoughneigh, 2026-10-04
  • Run 128 SMs with neither extra power phases nor a limiter: swap in the 300 W curve from the 8 GB overclock VBIOS. Source: @scallion_9883, 2026-10-16
  • Undervolt first; beyond that, a shunt mod with the missing power phases fitted. Source: @wakeupmbrc1, 2026-10-16
  • The no-flash VBIOS loader raised TDP to 300 W on drivers 610 and 615 and works with the Gen3 unlock. Source: @lxtend, 2026-10-20
• NVLink (NVLink). Not public, not even for 10 GB. Source: @lia_oc, 2026-10-22 One member says it works with a very simple hardware mod that is not his to share. Source: @bigptoughneigh, 2026-10-23 The missing parts are power phases and the NVLink clock circuit; Source: @globalapathy, 2026-10-24 the U11 clock generator (GS7612S5MA-R) is not fitted. Source: @onemillionkates, 2026-10-05 Compatible bridges: PNY RTXA6000NVLINK-KIT, PNY RTXA6000NVLINK3S-KIT, NVIDIA 900-53651-0000-000. Source: @djpandaman, 2026-10-06 NVLink on PCIe A100-class cards links pairs only. Source: @endsjeremy, 2026-10-08

What it is good for

• The 10 GB card is the better buy for training and image or video generation. Source: @wakeupmbrc1, 2026-10-13 For LLMs, VRAM is the bottleneck for most users, so trading 8 GB cards for 10 GB cards is not worth it. Source: @lia_oc, 2026-10-22
• Few recipes are written for 40 GB cards yet. Mountains started collecting them. Source: @mountains1605, 2026-10-10 Examples: Qwen 3.8 Flash Next (iq3_s) with Strata on one 40 GB card, where moving from x4 to x16 mainly improved prefill; Source: @exile_1ck, 2026-10-04 GLM 5.3 Flash 4-bit EXL3 across two 64 GB cards, one 40 GB card and an RTX 3080 20GB. Source: @vektor_prime, 2026-10-14 See All recipes and filter by 10 GB. Feralzi runs Nanbeige4.3-33B (dealignai's uncensored Q6_K GGUF) with about 200K context on a single 40 GB card at 27 tok/s while still tuning; see Recipes#10 GB card. Source: @feralzi, 2026-10-26
  Source: @feralzi, 2026-10-26
• Measured peaks at 122 SMs (clpeak; stock power settings by the owner's account, not double-checked): FP32 21.9 TFLOPS, FP16 tensor 312 TFLOPS, INT8 tensor 605 TOPS, cuBLASLt BF16 GEMM 308 TFLOPS. Full table: Benchmarks#Compute and memory peaks. Source: @djpandaman, 2026-10-23
  Source: @djpandaman, 2026-10-23
• Mixing a 40 GB and a 64 GB card works with a per-card unlock profile; fine for PP over Gen2, too slow for TP. Source: @djpandaman, 2026-10-07
• Compute: one unlocked 10 GB card mines 250 TH at 250 W on Pearl, against about 165 TH for akkerstone's 8 GB card with the core locked at 1,170 MHz. Source: @lxtend, 2026-10-21
  Source: @akkerstone, 2026-10-21

Older results at 40 GB (July to October 2026)

Runs from before the RIR burn, on cards at the 40 GB profile. Most are single posts; engines and quants as the posters named them. More on All recipes, llama.cpp on CMP cards and Image, video, voice and training.

• Reported Qwen3.6-35B-A3B q4_k_p on one card (Far, 2026-07-19). Mainline llama.cpp; the numbers are only in a screenshot. arnold civardagezen judged it better than his RTX 5060 Ti, but not by the 3 to 4x that full HBM bandwidth should give; simba6959: memory bandwidth is not locked, and the limits are PCIe and the SM count (70 at the time). Source: @a.civardagezen, 2026-07-19
  Source: @farawaythrower, 2026-07-19
  Source: @simba6959, 2026-07-19
• Reported GLM-5.2 Q4 with partial offload on 3x 10 GB (Eksitus). unsloth GLM-5.2 UD-Q4_K_XL (435 GB) in llama.cpp, split-mode layer across three cards at 40 GB, the rest on CPU and DDR4 (2x EPYC 7713). At Gen1 x4 (2026-07-20): pp2048 33.44 tok/s and tg512 5.90 tok/s; against CPU alone, generation rose about 60% and prefill fell about 30%. At Gen2 x4 (2026-07-24): pp2048 48.22 and tg512 6.39 tok/s. He said 64 GB cards would suit a model this size better. Source: @eksitus, 2026-07-20
  Source: @eksitus, 2026-07-24
• Reported Qwen3.6-27B on one card (Eksitus, 2026-07-22). unsloth Qwen3.6-27B-MTP-GGUF on one card at 40 GB, Gen1 x4, 250 W limit. UD-Q4_K_XL on ik_llama.cpp: tg128 33.2 tok/s and pp2048 545 tok/s without MTP, 41.1 tok/s decode with MTP (n_max=2); stock llama.cpp with draft-mtp: tg128 46.2 tok/s (peaks 55 to 59). UD-Q8_K_XL on ik_llama.cpp with MTP: about 38 tok/s. Every run had error bars of a few tok/s. Source: @eksitus, 2026-07-22
• Reported Six cards, Qwen 27B and 35B at 4-bit (bigfoot, 2026-07-24). All six of his 10 GB cards ran the models after the unlock, with no crash and no bricked card; without a good cooler they overheated after about 10 minutes. Source: @bigfoot1144, 2026-07-24
• Reported One vLLM model per GPU on six cards (vernongreen, 2026-07-24). GPUs 0 to 4 each served a vLLM OpenAI-compatible endpoint (Qwen2.5-7B, Qwen3-Coder-30B-A3B, Qwen2.5-32B, Qwen2.5-VL-32B and Qwen2.5-Omni-7B, all AWQ 4-bit) and GPU 5 ran ComfyUI: a concurrency smoke test at Gen1 with no throughput numbers. Source: @vernongreen, 2026-07-24
• Reported SGLang on one 40 GB card. makesier relayed a test of Qwen 27B FP8 on a single card under SGLang at about 1,800 tok/s prefill and 60 tok/s decode without MTP (2026-07-31). Ciender's benchmark repo runs Qwen3.6-27B W8A8 on SGLang 0.5.16 on a 40 GiB card and a 64 GiB card; it reports the 10 GB card going from 41.62 to 90.26 tok/s with MTP at 1024 in / 8192 out (acceptance length 2.91 to 3.50, against 4.00 on the 8 GB card). It is not a controlled A/B. Luna noted that the 40 GB configuration runs the VRAM notably slower. Source: @makesier, 2026-07-31
  Source: @lunaroseoc, 2026-08-01
  Source: repo github.com/Ciender/CMP-170HX-SGLang-Qwen3.6-Benchmarks by Ciender on GitHub, posted by @yuiayanami, 2026-08-01
• Reported Laguna-S-2.1 NVFP4 GGUF on 2x 10 GB (2026-08-11). Reported in a thread between bigfoot and You Know Who: a custom llama-server backend (from ciruinference) runs jcbtc/Laguna-S-2.1-NVFP4-GGUF at 131072 context with --parallel 30 --tensor-split 1,1 and a q8_0 KV cache, at about 53 tok/s decode. NVFP4 runs on Ampere through a fallback path. Source: @bigfoot1144, 2026-08-11
  Source: @squishysquare, 2026-08-11
• Reported Pruned DeepSeek V4 Flash on 4x 10 GB (cuddylac997, 2026-08-12). DeepSeek V4 Flash with 25% of its experts removed (REAP-style), on four cards at about 37 GB usable each (148 GB), Gen2 x4, not tensor parallel: 48.2 and "almost 50" tok/s. A 16-prompt test gave 11 pass, 1 fail and 4 to review, at 42.4 tok/s aggregate. Source: @cuddylac997, 2026-08-12
• Reported Qwen 3.8 27B Q6 with MTP on one card (Harmacist, 2026-08-20). 55 tok/s decode; engine, flags and context were not stated. Source: @romangitz, 2026-08-20
• Proven MiniMax H3 video on one card in an x1 slot (Lycorascence, 2026-08-26). After he patched it so it no longer overwrites its buffer and crashes the card, H3 at 16:9 0.5 MP takes 5 to 10 minutes for draft or moderate step counts with int8 convrot. At 1 MP an x1 card at 40 GB and a lane-restricted RTX 3090 run about even on s/it; Dino saw the same, with the CMP a couple of seconds ahead. Source: @lycorascence, 2026-08-26
  Source: @dino_38975, 2026-08-29
• Reported Qwen3.8-Flash-Next IQ4_XS on 2x 10 GB (2026-10-01). squishysquare: 550 tok/s prefill and 54 tok/s decode, with 48 GB of system RAM ("okay ish"). Source: @squishysquare, 2026-10-01
• Proven Supermicro SYS-1028GQ-TRT with 4x 10 GB at Gen2. cuddylac997's server (X10DGQ board, two Xeon E5-2650 v3, 256 GB DDR4) runs four 10 GB cards at 40 GB each (160 GB pooled, about 37.7 GiB usable per card), each at PCIe Gen2 x4. Idle is about 39 to 42 W per card at a 250 W limit; under gpu-burn -tc the cards reach 216 to 263 W, so the chassis applies no power-brake throttling. The 1U chassis holds four double-width cards lying flat and cost him $179 on eBay; programmerboi runs the same chassis (his prototype unit cost about $190). Humvee55 learned after buying that the TRT ships with an X9 riser board and the TRT2 with an X10 riser. See Supermicro SYS-1028GQ-TRT. Source: @cuddylac997, 2026-08-11
  Source: @scallion_9883, 2026-08-13
  Source: @cuddylac997, 2026-08-13
  Source: @programmerboi, 2026-08-23
• Reported Six 10 GB cards for rent on Vast.ai. RedPandaMining listed a 6x 10 GB rig at the 40 GB unlock on Vast.ai on 2026-08-29. Vast hides hosts with reliability below 90 by default; slide the host reliability filter fully left to find such rigs (Dino). Source: @redpandamining, 2026-08-13
  Source: @dino_38975, 2026-08-29

Known issues

Details: Troubleshooting, Xid errors, Driver and unlock failures.

• Warning Overheats in minutes without airflow. A new card clocked out in under 5 minutes with no cooling and showed in lspci but not in nvidia-smi. Source: @programmerboi, 2026-10-25 See Cooling and power.
• Dead on arrival: VBIOS ROM reads 0xFFFF. GSP init fails with "did not find valid ROM signature"; suspected bad SPI flash, reflash with a hardware programmer. Source: @youngju10000, 2026-10-08 The same error signature hit another member's card on bare metal on 2026-08-05, after an attempt at soldering capacitors; no fix was posted. Source: @costak, 2026-08-05
• Dead GPC blocks boot. On a card whose software-override gate is fused shut, a bad GPC6 blocks firmware boot, which blocks the PLM unlock, which blocks the RIR burn that would fix it. Unresolved. Source: @exile_1ck, 2026-10-11
• One card stays at Gen2 in a multi-card box. lxwx traced it to a physically damaged on-die PCIe controller on one card; the other three ran Gen3 together. Source: @lxtend, 2026-10-19
• Link retraining broke on kernel 7.2.11. Fixed by rolling back the kernel and NVIDIA driver after the RIR burn. Source: @lycorascence, 2026-10-23
• Reported VBIOS limits: memory fixed at 1215 MHz, no core offset, 250 W. After unlocking to 40 GB, GMZombie007 found that the lock on core clock offset and memory clock was still in place on 10 GB cards only (2026-07-19); Mr Pec had noted that the 8 GB card's VBIOS allows a higher memory clock and a core offset, which the 10 GB card's lacks. MaitOPS (2026-08-27): a 10 GB card is 250 W only. For a core offset, see the V/F curve item below. Source: @gmzombie, 2026-07-19
  Source: @mrpec, 2026-07-19
  Source: @_maitops, 2026-08-27
• Is there a 300 W VBIOS for the 10 GB card? Contested Nothing here was measured. The approaches people use to raise the limit are under #Limits. In the order posted:
  • btakk_ak47: 10 GB owners can flash a VBIOS that enables 300 W (TechPowerUp VGA BIOS entry 268495); no measurements were posted. Source: @btakk_ak47, 2026-07-24
  • SealClubber, answering Xeax: a 300 W BIOS on the 10 GB card is not possible. Source: @xeax, 2026-08-12
    Source: @sealclubber, 2026-08-12
  • MaitOPS and Lycorascence, both 10 GB owners: no 300 W VBIOS exists for the 10 GB card. Lycorascence says the stock limit is 250 W (transients aside) and he touches the VBIOS only for the memory unlock. Source: @lycorascence, 2026-08-25
    Source: @_maitops, 2026-08-25
  • Zagitta: per the Consensus-Protocol docs, the 300 W VBIOS exists only for the 8 GB card. Source: @zagitta, 2026-10-12
• Reported Undervolt with the 8 GB card's V/F curve. Big Ptoughneigh suggested giving the 10 GB card the 8 GB card's voltage/frequency curve, which enables a core clock offset as on the 8 GB card. eXile tried it the same day (2026-10-03): decode was 5%, 6%, 8% and 11% faster than stock at 4K, 32K, 128K and 261K prompt lengths, with average power slightly lower (157 W against 162 W). Idle still does not drop much below 25 to 30 W. See Undervolting and power limits. Source: @exile_1ck, 2026-10-03
  Source: @bigptoughneigh, 2026-10-03

  RH-3D says his V/F-type mod supports the 10 GB card too. Source: @rh3dai, 2026-10-03 Pointer: a longer write-up by @rh3dai exists; see Discord.

• Reported 170tune on a 10 GB card in a Dell R740. Before 170tune's 10 GB fix, nvml_oc on Pasteyy.'s card reported the GPC and memory V/F offset range as [0 .. +0]. After the fix, clock locking worked only up to 1410 MHz (1300 MHz also worked), and he doubted the core offset was being applied. See 170tune and Dell PowerEdge R740. Source: @pasteyy., 2026-07-31

Buying

Details: Buying.

• Dump the fuses before paying for a used 10 GB card: cards with wasted RIR records and no unlock are expected on the used market. Source: @dottoreargus, 2026-10-20
• Reported prices: about 200 USD per card before the unlock news; Source: @cuddylac997, 2026-10-07 all 10 GB cards had vanished from listings by 2026-10-07; Source: @kellertreppe123, 2026-10-07 in China the 10 GB card cost about 20,000 RMB on 2026-10-13, more than an 8 GB card, because it is a fully unlockable GA100; Source: @wakeupmbrc1, 2026-10-13 one sold "for parts" on eBay for 1,200 USD worked and unlocked to 40 GB. Source: @programmerboi, 2026-10-25
• Reported Memory bins decide the unlock. A pre-purchase research digest in the lab notes (2026-10-12; a summary of public sources, not lab measurements): the Samsung 10de:2082 bin unlocks to 40 GB only (Xid 79 beyond it), the Hynix 10de:20c2 bin to 64 GB. Buy one card and bench it before you buy four. Source: @darkmaniac7, 2026-10-12 (lab notes)
• Reported Mixed lots and 8 GB prices for 10 GB cards. On 2026-08-11 Schaka was offered a lot of eight cards (four 8 GB, four 10 GB) sold only as a whole and suspected the seller was offloading the weaker 10 GB cards. gpulover: 10 GB sellers asking 8 GB prices is ridiculous, since the 10 GB card stays at 40 GB usable. Source: @schakaqt, 2026-08-11
  Source: @artistosu, 2026-08-11
• Priced above or below the 8 GB card? Contested The direction flipped at least twice between July and October 2026; dated series on Price history and market and Price history, July to October 2026. In the order posted:
  • July, scarce and quoted above the 8 GB card: lxwx saw about 8000 CNY on Xianyu against 4000 for the 8 GB card; Humvee55 asked $1000 against $800 in the US; Eksitus's one quote had the 10 GB card $200 higher; arnold civardagezen put the premium down to sellers believing the 80 GB unlock works. Source: @lxtend, 2026-07-20
    Source: @scallion_9883, 2026-07-20
    Source: @eksitus, 2026-07-20
    Source: @a.civardagezen, 2026-07-20
  • Late July, the 8 GB card valued higher because it unlocks to 64 GB: quark2world's fair prices were $650 for the 8 GB and $600 for the 10 GB; machinezero put $800 against $600; diabeetee: "8gb cards are where its at". roger_dammit saw 10 GB cards $300 to $500 cheaper on eBay. Source: @quark2world, 2026-07-23
    Source: @machinezer0, 2026-07-26
    Source: @diabeetee, 2026-07-26
    Source: @roger_dammit, 2026-07-31
  • August, back above in places: DaLZy said Chinese sellers asked about $300 more for the 10 GB card; giaki3003 saw it sell for more than the 8 GB where he is, and segmond thought 80 GB rumours pushed prices up; on 2026-08-27 philow saw a single 10 GB card sell for 2500 EUR while orrorin saw 10 GB cards sit unsold on eBay at $1500. Source: @dalzyyy, 2026-08-10
    Source: @giaki3003, 2026-08-16
    Source: @segmond, 2026-08-16
    Source: @.philow, 2026-08-27
    Source: @_orrorin, 2026-08-27
  • September, below and then climbing: Humvee55 sold all his 8 GB cards on eBay but only three 10 GB ("no one wants 10gbs"); Colo: the 10 GB card tends to be a few hundred cheaper than the 8 GB; gym: anything cheap is almost certainly a scam, the 10 GB card is still a minimum of 1k; mrpandagod: 10 GB cards start around $1,800 on US eBay; OB88594 saw abnormal sales and price rises in China on 2026-09-22; by 2026-10-01 CUTE said a listing cost more than an A100 40G and SKYY said Chinese 10 GB prices had doubled. Source: @scallion_9883, 2026-09-09
    Source: @colo2442, 2026-09-10
    Source: @jimtjames, 2026-09-13
    Source: @djpandaman, 2026-09-14
    Source: @ob88594, 2026-09-22
    Source: @cute_sc2, 2026-10-01
    Source: @skyy2333, 2026-10-01
  • October, disputed again: lenxtroz put the 10 GB card $500 to $800 below the 8 GB, Zoodasa said it is priced well above because of its extra unlocks, and Miséricorde saw the two at about the same price on Goofish with the 10 GB still cheaper on eBay (2026-10-13). Feralzi found nothing under $2.5k on eBay and mrpandagod's sold data showed about $2 to 2.3k (2026-10-17); on 2026-10-18 akkerstone saw nobody selling. None of the 2026-10-13 posts gave completed-sale data. Source: @lenxtroz, 2026-10-13
    Source: @endbr64, 2026-10-13
    Source: @wakeupmbrc1, 2026-10-13
    Source: @feralzi, 2026-10-17
    Source: @djpandaman, 2026-10-17
    Source: @akkerstone, 2026-10-18
• Paid unlock services in China. Contested Hearsay, not verified on the Discord. A third party was said to sell a permanent 10 GB unlock for about USD 500 (Chengdu), pitched as more than an A100 40G PCIe, with NVLink extra. Source: @ob88594, 2026-10-09
  Source: @ob88594, 2026-10-09
  Source: @ob88594, 2026-10-09 The card then identifies as an A100 with no driver patches; Miséricorde reads it as an RIR macro burn. Source: @wakeupmbrc1, 2026-10-09 The unlocking merchants on Xianyu were swamped and took the service down. Source: @ob88594, 2026-10-12 The Xianyu seller offering a permanent unlock to A100 says only a few cards can reach 48 GB, and days before launching the service he had asked lxwx to sell his 10 GB cards back at market price. Source: @lxtend, 2026-10-17
  Source: @lxtend, 2026-10-17 Be careful with private unlock claims: a screenshot is easier to fake than a result when money is involved, and nobody has NVIDIA's production keys. Source: @djpandaman, 2026-10-17

See also

8 GB card · Variant comparison · CMP 170HX · Unlocks · Platforms · Glossary

Imported from the cmpunlocker Discord with credit; see Credits.