32 GB cards (subsystem 10de:1586)
More actions
Some cards sold as 8 GB CMP 170HX cards unlock to 32 GB, not 64 GB. They report the 8 GB card's device ID 10de:20c2, but PCI subsystem 10de:1586, board number 699-11001-0109-RC1 and VBIOS 92.00.6D.00.01; a normal 8 GB card reads subsystem 10de:1585 and board 900-11001-0108-000. They were first documented on 2026-08-31, several buyers lost money on them, and sealed, new-looking stock is no protection. Ask the seller for the subsystem ID, board number and VBIOS version before paying, buy where returns are possible, and check the unlocked capacity on arrival (Verifying an unlock). The other variants are on Which card do I have and Odd variants and engineering samples.
How to spot one
Read the three identifiers before you pay (ask the seller for the output) and again on arrival:
lspci -nnv -d 10de: | grep -i -E 'controller|subsystem'
nvidia-smi -q | grep -i -E 'board part number|vbios'
| Normal 8 GB card | 32 GB card | |
|---|---|---|
| Device ID | 10de:20c2 | 10de:20c2 on the cards documented so far (see #Device ID question) |
| Subsystem | 10de:1585 | 10de:1586 |
| Board part number | 900-11001-0108-000 | 699-11001-0109-RC1 |
| VBIOS | 92.00.67.00.01 or 92.00.6D.00.0A | 92.00.6D.00.01 |
| Unlocks to | 64 GB | 32 GB |
Subsystem 1586, board 699-11001-0109-RC1 and VBIOS 92.00.6D.00.01 mark a 32 GB card
Asked on 2026-09-07 whether an 8 GB card that unlocks to 64 GB can be told apart from one that stops at 32 GB, tony said yes: the 32 GB cards report PCI subsystem 10de:1586, board number 699-11001-0109-RC1 and VBIOS 92.00.6D.00.01, while a normal 8 GB card reports subsystem 10de:1585 and board 900-11001-0108-000. B-st's card (2026-09-11) was labelled GPU GA100-105D-A1, assy 699-11001-0109-RC1 and PB-30290. On 2026-09-22 Arcanum said the 32 GB boards run a different, incompatible VBIOS: the normal 170HX versions are 92.00.67.00.01 or 92.00.6D.00.0A and the 32 GB boards report 92.00.6D.00.01 ("afaik"); he asked anyone who has seen other versions to share them. douya's card (2026-10-13) read PCI 10de:20c2, subsystem 10de:1586, GPU 20C2-105-A1, board 699-11001-0109-RC1 and ES VBIOS 92.00.6D.00.01. Zeferrum: a production card has Board Part Number 900-11001-0108-000, an engineering-sample card 699-xxxxx-xxxx-RCx. On 2026-10-12 Zeferrum gave the 32 GB subsystem as 1586 and dottoreargus thought it most likely 1585; in every earlier report on this page the 32 GB cards read 1586 and normal 8 GB cards 1585, so 1585 most likely describes a normal card, and djpandaman has never seen a 32 GB card among his. On arrival, check the board number, subsystem and VBIOS, then look for missing SMD capacitors and dried pads.
Warning 8 GB
Source: @tony_smartecom, 2026-09-07
Source: @cibernox, 2026-09-07
Source: @boz._., 2026-09-11
Source: @humandrone8721_60165, 2026-09-22
Source: @_arcanum_, 2026-09-22
Source: @zefrench6961, 2026-10-12
Source: @dottoreargus, 2026-10-12
Source: @djpandaman, 2026-10-12
Source: @douya4305, 2026-10-13
Source: @darkmaniac7, 2026-10-19 (lab notes)
Treat a cheap or vague 8 GB listing as a possible 32 GB card
Wayz asked for the VBIOS version of these cards to help identify them (2026-09-03). matstatman ("Best case you get 32gb version"), gpulover ("32G version risk") and Beam treat a 32 GB card as a standing risk on cheap or vague listings, and gpulover warned that any listing where a 64 GB unlock is not guaranteed may reach only 32 GB. For a US eBay listing of 8 GB cards at $2500 (2026-09-20), Arcanum's advice was to get the seller to show the VBIOS version before buying. Check the board or assy number and the VBIOS version before paying, not after.
Warning 8 GB
Source: @wayz2u, 2026-09-03
Source: @artistosu, 2026-09-11
Source: @matstatman, 2026-09-20
Source: @artistosu, 2026-09-20
Source: @beamreaching1, 2026-09-20
Source: @segmond, 2026-09-20
Source: @_arcanum_, 2026-09-20
Source: @matstatman, 2026-10-12
Source: @shudza, 2026-10-12
"Brand new, never inserted" 170HX cards from China unlock to 32 GB
Sellers in China advertise "brand new, never inserted" 170HX cards (a photo posted by Zoodasa, also sent to B-st). Arcanum and B-st: these are PG199 boards reconfigured to 170HX specs, and they unlock only to 32 GB; douya paid 9,000+ yuan for one (October 2026). A $1,750 listing at a time when the 8 GB buyout was 13k RMB was suspected to be ES stock, and Xi'an Qujiang Reicha Microelectronics tried to sell B-st 32 GB stock. Miséricorde: NVIDIA, not the seller, set that device ID, but the seller knew, and the card is not worth 13 to 14k RMB because VRAM value scales logarithmically. tony (a seller) countered that 9 to 10k is reasonable for new cards "better than 3090". A new, sealed look proves nothing: read the identifiers above before paying.
Warning 8 GB
Source: @_arcanum_, 2026-10-13
Source: @boz._., 2026-10-13
Source: @douya4305, 2026-10-13
Source: @wakeupmbrc1, 2026-10-13
Source: @endbr64, 2026-10-13
Source: @tony_smartecom, 2026-10-13
Source: @rrickk420, 2026-10-13
Source: @chris0chris0chris, 2026-10-13
ffa's study: 32 GB is stable, 64 GB fails past 32 GiB
On 2026-08-31 ffa documented a card that identifies as device 20C2, the 8 GB card's ID, but carries board number 699-11001-0109-RC1, PCI subsystem 10de:1586 and VBIOS 92.00.6D.00.01, with a 1458 MHz stock memory clock. Configured for 32 GB it is rock solid (31 GiB verified twice). Configured for 64 GB it enumerates 64 GB but fails past 32 GiB with CUDA_ERROR_ILLEGAL_ADDRESS, and at 39 GiB it raises Xid 1 (GSP task exception: load access fault), which needs a function-level reset (PF FLR) to recover; the standard 64 GB profile fails with Booter error 0x56. Measured bandwidth was 1277 GiB/s at 1458 MHz against 1496 GiB/s at 1728 MHz on a normal 8 GB card (subsystem 1585, board 900-11001-0108-000), which ffa read as the same bus width, suggesting 4-Hi instead of 8-Hi stacks rather than a missing stack. tony tried every configuration on his, confirmed it is a 32 GB card and returned several. Humvee55 said the memory is single rank, and that setting double rank only aliases 64 GB into L2.
Proven 8 GB
Source: @24528jjkk_12439, 2026-08-31
Source: @tony_smartecom, 2026-08-31
Source: @scallion_9883, 2026-08-31
A longer write-up by @24528jjkk_12439 (ffa), with the register values for each memory configuration and the bandwidth runs, exists; see Discord. (Pointer: the full notes wait for the author.)
What they are
Repurposed A100 DRIVE (PG199) boards, cleared stock or rehoused SXM parts
Nobody has a definitive answer. The views, in the order posted:
- tony is "99% sure" the 32 GB cards are repurposed A100 DRIVE 32G (PG199) parts, which he says the NVIDIA driver already knows about; asked whether someone reballed them, he said "Nvidia did". Arcanum noted that the only official 32 GB parts are PG199 DRIVE units, which were SXM only, and quoted PG199 at about 13k RMB (tony: 11k, pulled ones 8.5k). Rickcal suggested an SXM part rehoused in a CMP shell; DaLZy doubts anyone has the facilities, and insomniac noted that the HBM is on-package, so it is factory work. DaLZy and Studebaker think NVIDIA cleared old GA100/PG199 stock as 170HX. bayley_w raised an unresolved puzzle: the clocks indicate HBM2e, but there were no 8 GB HBM2e stacks. Source: @tony_smartecom, 2026-09-02
Source: @_arcanum_, 2026-09-02
Source: @rrickk420, 2026-09-02
Source: @dalzyyy, 2026-09-02
Source: @_studebaker_, 2026-09-02
Source: @bayley_w, 2026-09-02
Source: @ins_mniac, 2026-09-02 - giaki3003 (2026-09-11) speculated that someone in China fitted 32 GB A100 parts into 170HX shrouds, or that these are a legitimate 32 GB variant; gpulover called them the "170HX Mini" (32 GB unlockable). Source: @boz._., 2026-09-11
Source: @ins_mniac, 2026-09-11
Source: @giaki3003, 2026-09-11
Source: @artistosu, 2026-09-11
Source: @jjb01957, 2026-09-11
Source: @robbbbbb7973, 2026-09-11 - _L_ said on 2026-09-21 that the 32 GB cards sold as CMP 170HX are basically a PG199 DRIVE on a 170HX PCB, that probably only a few thousand exist, that nobody noticed in the mining days because every card exposed 8 GB, and to search for 10de:1586. cibernox expected these cards to still have about 2.16 to 2.2 TB/s, which does not match ffa's measured 1277 GiB/s. Petri Krohn asked owners whether the memory is Hynix or Samsung, to photograph the strap resistors and to upload the VBIOS to TechPowerUp. Source: @lia_oc, 2026-09-21
Source: @darcidride, 2026-09-21
Source: @cibernox, 2026-09-21
Source: @petrikrohn_26497, 2026-09-21
Source: @_arcanum_, 2026-09-21 - On 2026-09-22 HumanDrone8721 linked Amogh's GA100 VBIOS comparison (CMP 170HX against A100 PCIe against DRIVE A100), which assigns VBIOS 92.00.6D.00.01 to a 32 GB A100 rather than to a CMP 170HX. Arcanum said this fits, since tony received an A100 DRIVE board that was configured and sold as a 170HX 8G, and quark2world added that these are probably PG199 parts on an A100 PCB that were simply sold as "CMP". That thread concluded that there is no genuine 32 GB CMP 170HX, but that some A100 DRIVE boards have been sold as one. Source: @humandrone8721_60165, 2026-09-22
Source: @_arcanum_, 2026-09-22
Source: @quark2world, 2026-09-22
Source: @dalzyyy, 2026-09-22
Contested 8 GB
Theory: the 170HX boards are PG199 parts NVIDIA could not sell
DaLZy (2026-09-16) noted that a PCIe version of PG199 apparently exists and suggested "all those pg170hx are probably just pg199 they couldn't sell". Later in that chat PG199 is described as a 32 GB SXM2 part selling for about $1,500, with third-party SXM2 adapters available. This is an unverified idea about where the boards came from; compare the section above, where members identify the 32 GB cards as PG199 boards sold as 170HX.
Theory
Source: @dalzyyy, 2026-09-16
Device ID question
Does a 32 GB card report the CMP device ID?
The two statements, in the order posted:
- ffa's card (2026-08-31) identifies as device 20C2, the 8 GB CMP 170HX ID, and tony and _L_ tell these cards apart by the subsystem ID 10de:1586. Source: @24528jjkk_12439, 2026-08-31
Source: @tony_smartecom, 2026-08-31
Source: @scallion_9883, 2026-08-31
Source: @tony_smartecom, 2026-09-07
Source: @cibernox, 2026-09-07
Source: @lia_oc, 2026-09-21
Source: @darcidride, 2026-09-21
Source: @cibernox, 2026-09-21
Source: @petrikrohn_26497, 2026-09-21
Source: @_arcanum_, 2026-09-21 - quark2world said on 2026-09-22 that the PG199-derived boards "did not have the CMP PCI ID" and were simply sold as "CMP". Source: @humandrone8721_60165, 2026-09-22
Source: @_arcanum_, 2026-09-22
Source: @quark2world, 2026-09-22
Source: @dalzyyy, 2026-09-22
The two conflict unless quark2world meant the subsystem ID. lspci -nn output from a 32 GB card would settle it; douya's later identity dump (2026-10-13, above) also reads 10de:20c2.
Contested 8 GB
Buyers' losses and prices
tony's sealed batch from a Taobao seller unlocked only to 32 GB
tony bought 7 brand-new, still-wrapped 170HX cards (zero dust on the heatsink) from a Shenzhen Taobao seller he visited in person. They show 8 GB stock like a normal 170HX but unlocked only to 32 GB after a full night of trying (2026-09-02); the board part number is 699-11001-0109-RC1, with the same power limit as a normal 170HX but a different BIOS. He tested one card, left the rest sealed and returned them (Taobao accepted the return), and planned to negotiate a discount to 6-7k RMB on the opened one; ffa had to return some cards too. The salesperson said they had 50 (tony guesses 100-200), another buyer bought 50 and resold most, and DaLZy and Zagitta saw the same listing photos, so Zagitta cancelled his order. tony said on 2026-09-07 that he lost a week on these cards and that the selling company had held them for more than half a year, bought them cheaply before the unlocks and profited heavily. Other owners of 32 GB cards were turning up in #general by then; the supplier was not named. Before this nobody had mentioned a 32 GB variant, so buyers assumed every 8 GB card unlocks to 64 GB.
Warning 8 GB
Source: @24528jjkk_12439, 2026-08-31
Source: @tony_smartecom, 2026-08-31
Source: @scallion_9883, 2026-08-31
Source: @tony_smartecom, 2026-09-02
Source: @cmurrayis, 2026-09-02
Source: @dalzyyy, 2026-09-02
Source: @zagitta, 2026-09-02
Source: @tony_smartecom, 2026-09-07
Source: @cibernox, 2026-09-07
Sellers of 32 GB cards: local pickup only, or a refund and then double the price
tony said on 2026-09-02 that some sellers of the 32 GB cards now allow only local pickup, so a return is impossible. Matej expects many buyers to discover the 32 GB limit only after paying 2,000 USD or more. Zagitta's seller insisted the cards would work as 64 GB and said it could not open new products to show evidence; it then accepted a full refund and quoted the price doubled, at $2700. Buy where returns are possible and verify the unlocked capacity on arrival.
Warning 8 GB
Source: @zagitta, 2026-09-02
Source: @matej87, 2026-09-02
Source: @tony_smartecom, 2026-09-02
B-st paid $645, then the seller wrote "Expandable to 32GB"
B-st paid $645 (2026-09-11) for a card labelled GA100-105D-A1, assy 699-11001-0109-RC1. After the order the seller emailed that it was "Expandable to 32GB", the expansion to be done by the buyer through "technical means", and B-st later found that his card's photos exactly matched an earlier 32 GB sale. insomniac said tony had the same experience, 32 GB for the "109". jjb valued such cards at about 15% below the 10 GB card's price.
Warning 8 GB
Source: @boz._., 2026-09-11
Source: @ins_mniac, 2026-09-11
Source: @giaki3003, 2026-09-11
Source: @artistosu, 2026-09-11
Source: @jjb01957, 2026-09-11
Source: @robbbbbb7973, 2026-09-11
More buyers found 32 GB cards in September: "rugpulled", and 2 of 23 cards
On 2026-09-12 ST reported that a member had been "rugpulled" with a card that unlocked only to 32 GB instead of 64 GB; clee noted that this is still 4x the memory it was sold with. darcidride (Naturidride) found 2 cards that unlock only to 32 GB among 23 cards from 3 lots (2026-09-21), paid about 900 EUR each for them and planned to resell them. Unlock every card on arrival, while it can still be returned.
Warning 8 GB
Source: @stcord, 2026-09-12
Source: @c133, 2026-09-12
Source: @darcidride, 2026-09-21
Source: @cibernox, 2026-09-21
Source: @quark2world, 2026-09-21
Source: @deced_123, 2026-09-21
Source: @lia_oc, 2026-09-21
Source: @darcidride, 2026-09-21
Source: @cibernox, 2026-09-21
Source: @petrikrohn_26497, 2026-09-21
Source: @_arcanum_, 2026-09-21
Read listings closely: reused photos and misleading titles
On 2026-09-27 Adust spotted an eBay listing for a 10 GB card claiming PCIe 4.0, starting at $100; matstatman thought the photo looked wrong (the PCIe slot too far from the bracket). On 2026-09-29 Kaatchi posted an AliExpress listing titled "CMP170HX with Water cooling block" with $489 shipping; mrpandagod pointed out that it is only the water block, and insomniac said to read AliExpress listings very carefully, since this one did not lie but gave the wrong impression. youngchap has caught Alibaba scammers using photos of cards previously sold on eBay; compare B-st's card above, whose photos matched an earlier 32 GB sale.
Warning 8 GB and 10 GB
Source: @adusts, 2026-09-27
Source: @matstatman, 2026-09-27
Source: @stcord, 2026-09-27
Source: @catchyuser, 2026-09-29
Source: @_youngchap, 2026-09-29
Source: @djpandaman, 2026-09-29
Source: @ins_mniac, 2026-09-29
Unlock behaviour
On a subsystem 1586 card the driver unlock stops at 32 GB: that is the card, not the tool (see ffa's study above). The reverse also happened once: a tool update capped normal cards at 32 GB.
cmpunlocker 0.3 capped some normal 8 GB cards at 32 GB (fixed on master on 2026-09-10)
rgminer dev reported on 2026-09-09 that unlocker v0.3 misdetected an 8 GB card as an "ES 32 GB" card and unlocked only 32 GB, while earlier versions gave the same card a correct 64 GB; two of his cards that had unlocked to 64 GB and worked showed 32 GB after installing v0.3, and he suspected a detection trigger bug. He worked around it locally by disabling some checks and posted a bug report archive. Amogh identified the cause, reverted the change and said at 2026-09-10 02:09 UTC that it was fixed on master; the fix is to pull current master of cmpunlocker. Earlier, stoddartc1 (2026-09-07) saw both his cards (one 10 GB, one 8 GB) stay at 8192 MB after the 0.3 install.sh while 0.2 had unlocked them; whether that had the same cause was not stated. In another thread insomniac said on 2026-09-07 that the 2026-09-06 update had been retracted and advised reverting to the older unlock. If a card that unlocked to 64 GB before shows 32 GB after an update, read its subsystem ID first: 10de:1585 points at the tool, 10de:1586 at the card.
Reported 8 GB and 10 GB
Source: @ins_mniac, 2026-09-07
Source: @stoddartc1, 2026-09-07
Source: @_j3ns, 2026-09-07
Source: @roman13036, 2026-09-09
Source: @roman13036, 2026-09-09
Source: @_j3ns, 2026-09-09
Files
- Stock VBIOS of a 32 GB card (pointer). A dump of the stock VBIOS of a 32 GB card (
cmp170hx-1586-92006D0001-stock.rom, version 92.00.6D.00.01) by @tony_smartecom (tony) exists; see Discord. It goes on the wiki once tony agrees. Source: @tony_smartecom, 2026-09-02