Which card do I have

Reported 8 GB and 10 GB

There are two production versions of the CMP 170HX: the 8 GB card and the 10 GB card. They look the same from the outside, use the same board and the same cooler, and both unlock with cmpunlocker. They do not unlock to the same thing. Find out which one you have before you buy parts, burn fuses or pick a recipe.

The short version: run lspci -nn. 10de:20c2 is the 8 GB card. 10de:2082 is the 10 GB card.

Quick answer

Check	8 GB card	10 GB card
PCI device ID (`lspci -nn`)	`10de:20c2`	`10de:2082`
PCI subsystem ID (`lspci -nnv`)	`10de:1585`	`10de:1557`
Memory before unlock	8 GB	10 GB
Memory after cmpunlocker	64 GB (65536 MiB)	40 GB (40960 MiB); 48 GB on some cards after more work
VBIOS versions seen	92.00.67.00.01, 92.00.6D.00.09, 92.00.6D.00.0A	92.00.66.00.02
Board part number (`nvidia-smi -q`)	900-11001-0108-000	900-11001-0105-000
GPU package marking	GA100-105F-A1	GA100-105A-A1
HBM vendor	SK hynix HBM2e	Samsung HBM2e (details)
Active memory partitions	16 of 24 FBPA, 4096-bit	20 of 24 FBPA, 5120-bit
Free RIR fuse records (factory)	none usable (15 of 16 used, last one partly used)	16 of 16 free
SMs after unlock	74 (public cmpunlocker v0.5)	74 public; up to 128 after a RIR burn, per card

Device and subsystem IDs, package marking and stock memory clocks are from the 170th Street specification page by Amogh (GitHub amoghmunikote). Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 The unlocked sizes are the geometry cmpunlocker programs for each PCI ID. Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub, posted by @rh3dai, 2026-10-08 VBIOS versions and board part numbers are from the Consensus-Protocol cmp170hx VBIOS inventory. Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @zagitta, 2026-10-12 The memory partition counts are from gpcprobe readouts. Source: repo github.com/matstatman/gpcprobe by matstatman on GitHub, posted by @matstatman, 2026-10-20

Look at the card

From the outside the two versions are the same card: same A100-derived PCB, same passive heatsink and shroud, same single EPS 8-pin power input, no display outputs. There is no reliable visual difference without software. Photos of both will be added here.

What you can check by eye:

The PCIe edge connector. A real 170HX board is out of the PCIe CEM spec: the gold fingers are missing the key notch on the left. A normal card would have to be cut down to fit a 170HX shell, so a swapped PCB inside a 170HX shroud is unlikely. Source: @dalzyyy, 2026-10-11
The bracket. Cards with the original bracket say "Made in China" on the back of it. Source: @spectakkular, 2026-10-04
The package. Both versions have a heat spreader (IHS) on the GPU package, like the A100 40 GB; the A100 80 GB does not. Source: @spectakkular, 2026-10-27 The package marking (GA100-105F-A1 on the 8 GB card, GA100-105A-A1 on the 10 GB card) is under the heatsink. Do not take the card apart only to read it; the software checks below are enough.
The memory vendor. If a label on the card or a tool shows the HBM vendor, SK hynix means the 8 GB card and Samsung the 10 GB card. See #Memory vendor.
Reseller stickers. Some cards carry an "aliiphone" sticker, the same one seen on eBay listings of AliExpress-sourced cards. Source: @lenxtroz, 2026-10-12
"Brand new" cards. There are no new 170HX cards on the market. "New, never inserted" means cleaned and relabelled ex-mining stock. Source: @chris0chris0chris, 2026-10-13 Listing photos of "new" cards are often the 32 GB engineering-sample boards described below. Source: @boz._., 2026-10-13

Check with lspci

This works on a stock card, before any driver or unlock is installed.

lspci -nn | grep -i 10de
lspci -nnv -d 10de: | grep -i -E 'controller|subsystem'

[10de:20c2] with subsystem 10de:1585: the 8 GB card.
[10de:2082] with subsystem 10de:1557: the 10 GB card. A probe of a 10 GB Samsung card read PCI device ID 0x2082. Source: @darkshad03, 2026-10-04
[10de:20c2] with subsystem 10de:1586: see engineering-sample card below.

The link line (LnkSta in lspci -vv) does not tell the variants apart. Both ship at PCIe Gen1 x4 and report a x16 slot capability. A seller screenshot that says "PCIe 4.0 x16" is reading the slot, not the card: the same screenshot showed the card at x1 Gen1 while idle. Source: @xeax, 2026-10-18

Check with nvidia-smi

nvidia-smi --query-gpu=name,pci.device_id,pci.sub_device_id,vbios_version,memory.total --format=csv
nvidia-smi -q | grep -i -E 'board part number|vbios'

pci.device_id reads 0x20C210DE on the 8 GB card (the same value 170tune prints in its preflight) and 0x208210DE on the 10 GB card.
VBIOS 92.00.6D.00.0A is 8 GB firmware only. Source: @port00, 2026-10-17 Two 8 GB VBIOS revisions share board part 900-11001-0108-000: 92.00.67.00.01 (250 W, HBM at NDIV 54, 1,458 MHz) and 92.00.6D.00.0A (300 W, NDIV 64, 1,728 MHz). Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 The 300 W VBIOS exists for the 8 GB card only. Source: @zagitta, 2026-10-12
Board part number 900-11001-0108-000 is the normal 8 GB card. An engineering-sample (32 GB) card shows a 699-xxxxx-xxxx-RCx board number instead. Source: @zefrench6961, 2026-10-19
Name. Do not rely on the name string. cmpunlocker carries a patch whose only job is to make the driver name the card "NVIDIA CMP 170HX"; it uses the same name for both variants. Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub, posted by @rh3dai, 2026-10-08
Memory after the unlock. cmpunlocker sets 65536 MiB on 20c2 cards and 40960 MiB on 2082 cards; its verify.sh treats 60000 MiB or more as a good 8 GB unlock and 35000-59999 MiB as a good 10 GB unlock. Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub, posted by @rh3dai, 2026-10-08 Eight unlocked 8 GB cards in one dump all reported 65536 MiB. Source: @newman_08462, 2026-10-09 Some tools report an unlocked 8 GB card as 65024 MiB with about 1,493 GB/s bandwidth. Source: @kio063218, 2026-10-16

Check the memory layout

gpcprobe is a read-only kernel module that reads the floorsweep fuses with the NVIDIA driver still loaded. It shows which memory partitions (FBP, FBPA) and GPCs are enabled, disabled or marked defective. Source: repo github.com/matstatman/gpcprobe by matstatman on GitHub, posted by @matstatman, 2026-10-20

8 GB card: 16 of 24 FBPA active (4096-bit bus). FBP 6 and FBP 11 are marked defective on almost every card; FBP 1 and 4 are only disabled. Source: @petrikrohn_26497, 2026-10-10 The unlock keeps the same 16 FBPA and widens each from a 512 MiB window to 4 GiB, which is where 64 GB comes from. Source: repo github.com/matstatman/gpcprobe by matstatman on GitHub, posted by @matstatman, 2026-10-20
10 GB card: 20 of 24 FBPA active (5120-bit bus), two FBPs off. All six 10 GB cards bigfoot tested had exactly two defective FBPs. Source: @bigfoot1144, 2026-10-09 A typical readout: "NVIDIA CMP 170HX 40 GiB", GPC healthy 5 / disabled 3 / defective 1, FBP healthy 10 / disabled 2 / defective 2. Source: @ghostinthedatacenter_87281, 2026-10-07 A few 10 GB cards show the sixth stack as disabled but not defective; those are the ones that can reach 48 GB. See 10 GB card: Memory.

Check the RIR fuse records (10 GB buyers)

The extra 10 GB unlocks (more SMs, PCIe Gen3) use a permanent RIR fuse burn. The 8 GB card arrives with its RIR records already used, so it has nothing to burn. A used 10 GB card may already have spent some. Warning Dump the fuses (gpcprobe or 170hx-rir-tool --status) before you pay for a used 10 GB card; people expect cards with wasted RIR records and no unlock to reach the used market. Source: @dottoreargus, 2026-10-20

8 GB card: 15 of 16 slots used, one factory record (0x0923), zero free. Two 8 GB cards dumped read-only were byte-identical. Source: @marklintonca, 2026-10-10 The 16th slot is not fresh either, it is partly burned. Source: @djpandaman, 2026-10-17
Unused 10 GB card: the tool shows "0/8 slots in use, 16 free" (two records per 32-bit row). Source: @programmerboi, 2026-10-25 One write is enough for the unlock. Source: @xeax, 2026-10-25

The burn itself is on RIR fuse burn and 10 GB card. Do not burn anything on an 8 GB card.

What each one unlocks to

	8 GB card	10 GB card
Memory	64 GB	40 GB; 48 GB if the sixth stack is healthy (about 1 card in 5)
SMs (stock 70)	74	up to 128, depends on the card (112, 122 and 128 seen)
PCIe speed (software)	Gen2	Gen2 public; Gen3 after the RIR burn; Gen4 reported, not public
PCIe width	x4 stock; x16 with the PCIe capacitor mod	same
NVLink	no path	not public; needs a board-level mod
ECC	yes (cmpunlocker v0.5)	yes (cmpunlocker v0.5)

The 8 GB card unlocks to 64 GB, 74 SMs and Gen2; the 10 GB card unlocks to 40 GB (48 GB with a good sixth stack) but gets 128 SMs, better interconnect and faster compute. Neither is strictly better: for LLMs memory matters most, for diffusion the 10 GB card is better. Source: @wakeupmbrc1, 2026-10-07
Proven The 8 GB card unlocks to 64 GB. Eight unlocked 8 GB cards in one dump all reported 65536 MiB, Source: @newman_08462, 2026-10-09 and 64 GB is the maximum confirmed on a production 8 GB card. Source: @djpandaman, 2026-10-12
Proven The 10 GB card unlocks to 40 GB, and most cards stop there. Source: @zorg33, 2026-10-07
Source: @wakeupmbrc1, 2026-10-07
Source: @programmerboi, 2026-10-25
Reported 48 GB needs a good sixth stack (a non-defective sixth FBP), about 1 card in 5: 6 of 30 cards mrpandagod saw tested could do it, Source: @djpandaman, 2026-10-17 and the estimate on the Discord is 20% or less. Source: @artistosu, 2026-10-26
The RIR path is 10 GB only: its RIR rows are not blown, so a software fuse remask can be enabled. Source: @lxtend, 2026-10-21
Source: @wakeupmbrc1, 2026-10-21
Source: @lxtend, 2026-10-21 On 8 GB cards most fuses are burnt and locked out. Source: @68x0, 2026-10-26
10 GB unlocked cards land at 112, 122 or 128 SMs. Source: @wakeupmbrc1, 2026-10-27

The full table with bandwidth, power and prices is on Variant comparison.

80 GB on the 10 GB card

Contested Not achieved. Some cards have read 80 GB, but the memory was unstable. Source: @tekkenz, 2026-10-12 The two readings, in the order they were posted:

The extra capacity is address aliasing ("folding"), not memory. Colo gets folds in every experiment, including physical accesses that bypass L2 and PRAMIN, unless a further unlock frees address bit 14. Source: @colo2442, 2026-10-14 An author of the earlier "80 GB" paper now also reads it as aliasing in L2. Source: @scallion_9883, 2026-10-16
The 10 GB stacks are physically 16 GB parts binned to half by Samsung, per a Chinese write-up ("170HX 10G unlock to 80G exploration", blog.kkk.rs/archives/71, title translated from Chinese); unlocking the upper half would need Samsung's internal vendor codes. Source: @djpandaman, 2026-10-17

Plan on 40 GB. Treat 48 GB as a lottery and 80 GB as research.

Memory vendor

Reported The 8 GB card uses SK hynix HBM2e and the 10 GB card uses Samsung HBM2e. Source: @darkmaniac7, 2026-10-30 (site owner, from his own cards) If a tool or a label on the card shows the memory vendor, that tells the versions apart: SK hynix is the 8 GB card, Samsung the 10 GB card.

Earlier posts that support or conflict with this, in the order posted:

Supporting: a probed 10 GB card was a Samsung part. Source: @darkshad03, 2026-10-04
Supporting: a read-only fuse study of one card of each kind read the 8 GB card as Hynix and the 10 GB card as Samsung. Source: repo github.com/ahnguyen17/ga100-fuse-re by ahnguyen17 on GitHub, posted by @exile_1ck, 2026-10-07
Supporting: two 8 GB cards read an SK hynix memory strap. Source: @marklintonca, 2026-10-10
Conflicting on the memory type: the 10 GB card has plain HBM2 on 5 active stacks; 16 GB-per-stack HBM2 never shipped. Source: @atsasin2126, 2026-10-16
Supporting the HBM2e type: the 10 GB stacks are binned 16 GB HBM2e, per the kkk.rs write-up. Source: @djpandaman, 2026-10-17
Conflicting on the memory type: the 10 GB card carries Samsung HBM2. Source: @rh3dai, 2026-10-17
Related: 80 or 96 GB is blocked by things not yet understood on the HBM side for the Samsung stacks. Source: @dalzyyy, 2026-10-18

Cards that are not a normal 170HX

Warning

32 GB engineering-sample boards (PG199 origin). These report the 8 GB device ID but unlock to 32 GB only. One bought for 9,000+ yuan identified as: vendor:device 10de:20c2, subsystem 10de:1586, GPU part 20C2-105-A1, board 699-11001-0109-RC1, ES VBIOS 92.00.6D.00.01. Source: @douya4305, 2026-10-13 They are PG199 boards reconfigured to 170HX specs. Source: @_arcanum_, 2026-10-13 NVIDIA repurposed them; a seller cannot change the device ID. Source: @wakeupmbrc1, 2026-10-13 The subsystem ID on such cards was also reported as 1585, so check the board number and VBIOS too. Source: @dottoreargus, 2026-10-19 A large Taobao store (shop35050154) listed a lot of "170HX 8GB" cards at 9,999 RMB, just under USD 1,500; they were PG199HX-origin (Tesla) cards. Source: @spectakkular, 2026-10-11
Source: @dalzyyy, 2026-10-11 An Alibaba supplier, Lilian Lilian of Xi'an Qujiang New District Reicha Microelectronics, tried to sell a member 32 GB stock; another member warned not to buy from them. Source: @kellertreppe123, 2026-10-11
Source: @boz._., 2026-10-23
Single-rank 8 GB cards. A community report describes 8 GB cards that stop at 32 GB after unlock (single-rank) next to the normal dual-rank 64 GB cards; three of four cards in that report were dual-rank. A faked 64 GB unlock on such a card passes memtest_vulkan but fails a CUDA memtest, so validate with a CUDA-side test. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 Whether these are the same boards as the engineering samples above is not known.
Fakes and transplants. The reported overall failure rate for bought cards is about 20%. Fake cards have been reported, and one "CMP" turned out to carry a die transplanted from a 32 GB A100. Source: @civiloid, 2026-10-13
Other CMP cards. The CMP 90HX, 100-210 and the rest are different silicon. Do not confuse a 10 GB CMP 90HX with the 10 GB 170HX: an unlocked 90HX is 3090-class compute but keeps 10 GB of GDDR6X. Source: @javicabrera_30495, 2026-10-18 The Titan V is GV100 (Volta, like the V100), not GA100, so it is no relative of the 170HX; its mining counterpart is the CMP 100HX (100-210). This came up when comet1310 asked whether a Titan V can be unlocked to 16 GB. Source: @c07f, 2026-10-24
Source: @endsjeremy, 2026-10-24
Source: @endsjeremy, 2026-10-24
Source: @c07f, 2026-10-24 Not every one of those can be modded: there are different revisions and you have to test; some have been unlocked, but at V100 prices it is not very lucrative (xeax). Source: @xeax, 2026-10-24
Source: @xeax, 2026-10-24
Source: @xeax, 2026-10-24 A listing that says only "CMP 8GB" may not be a 170HX; check that it says 170HX. Source: @djpandaman, 2026-10-11 This wiki covers the 170HX.

If you are buying, see Buying and test before you pay.