Odd variants and engineering samples

Most cards sold as a CMP 170HX are one of the two production versions, the 8 GB card and the 10 GB card (Which card do I have). This page collects the rest: cards with fewer SMs than normal, engineering samples (ES) of the 170HX and of the A100, and single cards that did not behave. ES boards are pre-production cards that NVIDIA did not intend for sale; they can carry different board numbers, firmware and floorsweeping, and several did not unlock or did not boot. The 8 GB-looking boards that unlock only to 32 GB have their own page, 32 GB cards (subsystem 10de:1586).

Quick checks on arrival

SM count. Read the multiprocessor count, not the compute capability (sm_80 on every card). A normal card shows 70 before any SM unlock; see #Cards with 56 SMs. With PyTorch installed:

python3 -c "import torch; [print(i, torch.cuda.get_device_properties(i).multi_processor_count) for i in range(torch.cuda.device_count())]"

Board part number. nvidia-smi -q | grep -i "board part number" tells a production board from an ES board; see #Board part number tells an ES card from a production card.
Unlocked memory. Unlock the card and confirm the full size with a CUDA memtest while you can still return it: Verifying an unlock. An 8 GB card that stops at 32 GB belongs on 32 GB cards (subsystem 10de:1586).
Link and clocks. One owner returned a card that would not take Gen2 and would not clock past about 1200 MHz after the unlock; see #A 40 GB card that would not take Gen2 and stayed near 1200 MHz.

Cards with 56 SMs

Some cards have only 56 SMs instead of 70

Most cards have 70 SMs, but some report only 56 and are slower. Capone has several 8 GB cards with 56 SMs; they carry Aliphone stickers and he thinks they came from the Ohio charity sale in June 2026. Humvee55 owns a 10 GB card with only 56 active SMs: its memory stacks are fine and its SMs failed. Capone asked whether a fix for 56-SM cards had been published; none was mentioned. Rickcal speculated that such cards were unlocked and benchmarked in China first. The advice to buyers was to resell a 56-SM card, and RonBurgundy planned to buy extra cards to be sure of 70-SM ones. An estimate of about 67 SMs (5/8 of the A100's 108) was answered with the 70 that nvidia-smi shows; sm_80 is the CUDA compute capability, not the SM count, so check the multiprocessor count of every card on arrival.

Reported 8 GB and 10 GB
Source: @beefbonecapone, 2026-07-30
Source: @bigfoot1144, 2026-07-30
Source: @_arcanum_, 2026-07-30
Source: @rrickk420, 2026-07-30
Source: @ju5t5omedude, 2026-07-30
Source: @dalzyyy, 2026-07-31
Source: @cute_sc2, 2026-07-31
Source: @sandberg54, 2026-07-31
Source: @a.civardagezen, 2026-08-04
Source: @cute_sc2, 2026-08-04
Source: @ju5t5omedude, 2026-08-04
Source: @scallion_9883, 2026-08-04
Source: @beefbonecapone, 2026-08-04
Source: @programmerboi, 2026-08-04
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

Claim: 56-SM cards can be raised to 70

Capone claimed on 2026-07-31 that the lower-SM cards can go from 56 to 70 SMs because their GPCs are not disabled ("they aren't disabled GPC"). No method or proof was posted.

Theory 8 GB
Source: @beefbonecapone, 2026-07-31

Engineering-sample CMP cards

Board part number tells an ES card from a production card

On 2026-09-10 Amogh asked whether a problem card was an ES or a production card, since ES cards have a different floorsweep. Newman, who has tested an ES card, said the Board Part Number tells them apart: production 8 GB cards read 900-11001-0108-000 and ES cards 699-xxxxx-xxxx-RCx. Zeferrum restated the same check on 2026-10-12. Read it with:

nvidia-smi -q | grep -i "board part number"

The 32 GB boards read 699-11001-0109-RC1; see 32 GB cards (subsystem 10de:1586). Background on floorsweeping: Fuses and floorsweep maps.

Reported 8 GB
Source: @newman_08462, 2026-09-10
Source: @_j3ns, 2026-09-10
Source: @zefrench6961, 2026-10-12

A card labelled "not for sale, for development only"

Bagster, a seller shipping cards, found a card labelled "not for sale, for development only", with darker NVIDIA lettering and stickers reading GA100-105F-A1, ROM version 1001_0108_105.rom and model P1001 (P1001 is the A100 board number). The cards had probably all been flashed with mining ROMs, so the original VBIOS may be gone. He also posted a second variant with only a sticker and nothing on the bracket. DaLZy thought it is an ES or QS (qualification sample) card; Amogh said nothing special came to mind. Suggested tests: dump the MMIO registers with the probe script to compare memory fusing (xanthe_), check whether the card is production-fused (gk_hun), and dump the VBIOS in case it is not archived yet (Zeferrum). GA100-105F-A1 is also the package marking of a normal 8 GB card (Which card do I have#Quick answer), so the sticker alone does not identify an ES.

Reported 8 GB
Source: @bagstah, 2026-08-23
Source: @dalzyyy, 2026-08-23
Source: @_j3ns, 2026-08-23
Source: @xanthefin, 2026-08-23
Source: @gk_hun, 2026-08-23
Source: @zefrench6961, 2026-08-23

"170HX ES" cards sold as the official version cannot unlock to 64 GB

Warning 8 GB OB88594 reported on 2026-09-03 that a 170HX ES (engineering sample) version was being sold as the official version, and that it cannot be unlocked from 8 GB to 64 GB. The bracket is marked RC1 but may be swapped for a blank one, so the only sure check is to unlock the card on arrival and confirm 64 GB (Verifying an unlock). CUTE added that sellers had little stock left and that anyone without cards was in a bad position. RC1 fits the 699-xxxxx-xxxx-RCx board number of ES boards (above); the thread does not say whether these are the 32 GB boards.

Source: @ob88594, 2026-09-03
Source: @cute_sc2, 2026-09-03

Remote access to CMP 170HX engineering samples

On 2026-08-21 Amogh said he had got access to a couple of CMP 170HX engineering-sample cards through connections who let him use their machines remotely. He had already obtained most of what he needed from an ES (VBIOS, register map and so on). His goal is a VBIOS patched to run on a production card; he called the idea that ES cards accept any VBIOS "mostly a gimmick", and said production-fused A100s are enough to observe functional differences. He had not used drivers on the ES, and nvidia-smi did not work with it.

Reported
Source: @_j3ns, 2026-08-21
Source: @armand0e, 2026-08-21
Source: @bayley_w, 2026-08-21

An ES card with an NVIDIA fuse daughterboard

Humvee55 said on 2026-09-29 that the only fuse board he has seen is on Miranda's engineering-sample card. Miranda posted photos of the card with a blue NVIDIA-branded daughterboard attached and invited people to look at it ("anyone wanna poke at my monstrosity?"). It came up in a discussion of the theory that RIR or fuse programming on older generations needed a hardware daughterboard. Background: Fuses and floorsweep maps.

Reported
Source: @scallion_9883, 2026-09-29
Source: @mirandahw, 2026-09-29
Source: @_arcanum_, 2026-09-29

GeekBench identifies the 170HX as an engineering sample

Amogh said on 2026-07-21 that he once ran GeekBench on the 170HX and it detected the card as an engineering sample ("it thought it was an ES"), so the benchmark is not usable on it. A GeekBench ES label is therefore no test for an ES board; use the board part number above. Benchmarks that work: Benchmarks.

Reported
Source: @_j3ns, 2026-07-21

A100 engineering samples

Looking for a live A100 ES

On 2026-08-06 SealClubber linked an A100 engineering-sample VBIOS on TechPowerUp (A100 40960 MB, entry 277449). Amogh said the file helps but he needed a live A100 ES card to run scripts on; Samuel Hayden had not found anyone with one, and Jeremy (Ends) suggested asking Wendell of Level1Techs. On 2026-08-21, with the CMP ES cards available (above), Amogh deprioritized the A100 script. He pointed to an A100 ES on eBay for about $2000 with a board footprint similar to GA100 and a lot of extra GPIO, and said "On the VBIOS side, there are some juicy things there."

Reported
Source: @sealclubber, 2026-08-06
Source: @_j3ns, 2026-08-06
Source: @samuel_hayden., 2026-08-06
Source: @endsjeremy, 2026-08-06
Source: @_j3ns, 2026-08-21

The unlocker fails on an A100 ES1 at NVRM initialisation

K2V asked on 2026-07-22 whether anyone had succeeded on A100 engineering samples rather than the 170HX. He tried an A100 ES1 (20B0, 8192 MB, 2048-bit, 4096 CUDA cores, Samsung 8-Hi HBM2) and could not get past the NVRM initialisation error, even though the SEC2_DEBUG readout showed the registers written. No success was reported; the unlocker targets the 170HX.

Reported
Source: @k2v0008, 2026-07-22

An ES A100 bought on eBay did not boot

Warning An eBay development or ES A100 had been discussed as a possible source of knowledge about the 170HX, and bayley_w bought it. It did not boot on Windows or Linux, possibly because of bad firmware (2026-09-11). He had not yet pulled the firmware from the flash chip; firmware images for this SKU are around, so it may be recoverable. Dollop's takeaway: with ES cards it is hard to tell before buying whether you are getting a paperweight.

Source: @dollop1358, 2026-09-11
Source: @bayley_w, 2026-09-11

Other odd cards and markings

A 40 GB card that would not take Gen2 and stayed near 1200 MHz

orrorin had two 64 GB cards and one 170HX that unlocked to 40 GB (2026-09-03). The 40 GB card would not unlock to PCIe Gen2; an AI chatbot suggested it is fused differently. CUTE disagreed: it should work, and the unlocker probably just does not handle that variant well ("40gb didn't got that much love"); earlier he had said PG170HX and RC1 cards should also be unlockable to 64 GB, perhaps with the same problem as the 10 GB cards. The card passed memory tests but would only clock to about 1200 MHz, even after re-running the unlock and locking clocks at 1700 MHz with nvidia-smi, so orrorin returned it as unhealthy. 40 GB is the 10 GB card's unlocked size, but the thread does not give this card's device ID. Gen2 itself: Gen2 link retrain.

Reported
Source: @_orrorin, 2026-09-03
Source: @cute_sc2, 2026-09-03

GA100 die SKU may differ between cards (GA100-105A or GA100-105AAA)

During the 80 GB tests on 2026-09-30 (The 10 GB card: provenance and the 80 GB question) Mello Jello asked whether the cards were GA100-105A or GA100-105AAA, which can be read only from the die marking. Lycorascence asked whether a register could reveal it; nobody answered. Lycorascence advised testing on several cards because there are apparently multiple chip SKUs, and JackSmith suggested results may be a binning lottery. For comparison, a 32 GB card was marked GA100-105D-A1 (32 GB cards (subsystem 10de:1586)); the markings of the two production versions are in Which card do I have#Quick answer.

Theory 10 GB
Source: @mellojello_, 2026-09-30
Source: @lycorascence, 2026-09-30
Source: @jacksmith7981, 2026-09-30

Fused quality class: leaky part (1/4)

matstatman posted on 2026-09-26 a readout from his 170HX showing NVIDIA's fused quality classification of the GPU: "Indicate the exact quality level of GPU: leaky part (1/4)". The message did not name the tool or register.

Reported
Source: @matstatman, 2026-09-26

32 GB cards

Boards that report the 8 GB device ID (10de:20c2) but unlock only to 32 GB have their own page: 32 GB cards (subsystem 10de:1586). They read subsystem 10de:1586, board number 699-11001-0109-RC1 and VBIOS 92.00.6D.00.01. Source: @tony_smartecom, 2026-09-07
Source: @cibernox, 2026-09-07

Which card do I have, 32 GB cards (subsystem 10de:1586), Verifying an unlock, Acceptance checklist, Testing and burn-in, Full SM unlock, Fuses and floorsweep maps, RIR fuse burn, VBIOS and firmware, Gen2 link retrain, Benchmarks, Variant comparison, The 10 GB card: provenance and the 80 GB question, Other CMP cards and other NVIDIA GPUs, Card failure rates and longevity, Buying.