Jump to content
Toggle search
Toggle menu
CMP Wiki
Toggle preferences menu
Toggle personal menu
Not logged in
Your IP address will be publicly visible if you make any edits.

8 GB card

From CMP Wiki
More actions
Revision as of 08:03, 30 October 2026 by ImportBot (talk | contribs) (Import: everything for the 8 GB card, @wakeupmbrc1, @djpandaman, @c133, @ob88594, @mountains1605, @rh3dai, @matstatman, @replacedriver47, @reala7096, @alanoo, @mistyashhole, @lenxtroz)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

Reported 8 GB

The 8 GB card is the CMP 170HX with PCI ID 10de:20c2 and SK hynix HBM2e. It is the more common version and the one most recipes are written for. After the public unlock it is a 64 GB, 74 SM GA100 card on a PCIe Gen2 link. Not sure which card you have? See Which card do I have. New owner? Start with Start here#8gb.

At a glance

Stock After cmpunlocker v0.5 With hardware work
Memory 8 GB 64 GB (65536 MiB) same
SMs 70 74 same; no further path known
PCIe Gen1 x4 Gen2 x4 (Gen2 link retrain) Gen2 x16 with the PCIe capacitor mod
ECC off on (DRAM and SRAM) same
BAR1 64 MiB 64 GiB same
P2P none with the P2P patches (not on cmpunlocker master) same
NVLink fused off none no known path
Power limit 250 W or 300 W, depends on VBIOS same same

Unlock geometry, BAR1 and ECC per the cmpunlocker README and constants. Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub, posted by @rh3dai, 2026-10-08. 74 SMs measured with torch on four cards, Gen2 x16 on modded cards, 64 GiB BAR1 measured. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20

Unlock results

Details and steps: Driver unlock, cmpunlocker forks and Unlocks.

  • Proven The public state is cmpunlocker v0.5: 64 GB, +4 SMs (70 to 74) and ECC. RH-3D posted a walkthrough video of v0.5. Source: @rh3dai, 2026-10-11 v0.5 added +4 SMs for both variants; more SMs exist on the 10 GB card, not on the 8 GB. Source: @lia_oc, 2026-10-22 ECC is the other new piece. Source: @reala7096, 2026-10-22 Amogh has said there is not much more to get through software, so v0.5 is likely the last feature release. Source: @reala7096, 2026-10-22
  • One owner ran v0.5 with ECC for 25 h: 64 GB, memtest clean, 0 ECC errors. Source: @mistyashhole, 2026-10-10 A first-time owner: booted, installed the driver, ran the unlock, and it did the rest. Source: @lenxtroz, 2026-10-13
  • If your card shows 70 SMs after the unlock, you are on an older build; the current one gives 74. Source: @matstatman, 2026-10-08
  • The 8 GB unlock writes nothing permanent to the card. Only the special 10 GB unlocks burn fuses. Source: @ob88594, 2026-10-12 The unlocks are volatile: the patched driver reapplies them every boot and a cold power cycle returns the card to stock. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20
  • Rough edge: some 8 GB cards are single-rank and stop at 32 GB. Validate the unlock with a CUDA-side memtest, not only memtest_vulkan. See Which card do I have#Cards that are not a normal 170HX.

Memory

  • Layout. SK hynix HBM2e, 16 of 24 FBPA active (4096-bit): two full stacks and four half stacks. FBP 6 and 11 are physically dead; FBP 1 and 4 are only floorswept. The 64 GB unlock keeps the same 16 FBPA and widens each from a 512 MiB window to 4 GiB. Source: repo github.com/matstatman/gpcprobe by matstatman on GitHub, posted by @matstatman, 2026-10-20. Almost every 8 GB card shows FBP 6 and FBP 11 marked defective. Source: @petrikrohn_26497, 2026-10-10
  • Clock. Stock HBM clock depends on the VBIOS: 92.00.67.00.01 runs NDIV 54 (1,458 MHz), 92.00.6D.00.0A runs NDIV 64 (1,728 MHz), same board and timings. Raising two NDIV 54 cards to 64 with 170tune gave +5.8-6.3% single-user decode on GLM-5.3-Flash TP4. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20
  • Overclock. The HBM divider is set with 170tune. Source: @wakeupmbrc1, 2026-10-12 Hynix cards usually reach NDIV 68-70; 74 fails. Source: @alanoo, 2026-10-14 The gain depends on the workload: NDIV 54 to 70 gave +10.5% at c=1 on dense Qwen3.8-27B, Source: @wakeupmbrc1, 2026-10-14 NDIV 70 gave 3-4% on the same model for another owner, Source: @fnerdman, 2026-10-14 and alanoo measured zero gain up to NDIV 72 on MoE workloads. Source: @alanoo, 2026-10-14 See Glossary#HBM clock win.
  • Is the unlocked memory reliable? Contested No failure statistics exist yet; validate each card with a full memtest. The views, in the order posted:
    • It is lower-binned memory NVIDIA fused off for a reason, and software cannot fix that. Source: @_____josh_____, 2026-10-05
    • It runs fine and, apart from speed, no different from a 5090. Source: @jjb01957, 2026-10-05
    • Every card runs at NDIV 68 at least, better than typical binned memory. Source: @wakeupmbrc1, 2026-10-05
  • ECC does not change usable memory; HBM2e carries it natively. Source: @alanoo, 2026-10-05
  • Cards with HBM errors. A longer write-up by @zombielurker (Zombie) exists; see Discord. More on Troubleshooting.
  • 80 or 96 GB? Contested Not achieved. Screenshots of 8 GB cards at 80 or 96 GB circulated in October 2026 with nothing behind them. The positions, in the order posted:
    • A read-only fuse study puts the ceiling at 80 GiB (reviving FBP 1 and 4) and calls 96 GiB wrong because FBP 6 and 11 are dead silicon; its 80 GiB recipe is derived, incomplete and untested. Source: repo github.com/ahnguyen17/ga100-fuse-re by ahnguyen17 on GitHub, posted by @exile_1ck, 2026-10-07
    • In theory an 8 GB card could go to 96 GB if an unlock ever shipped. Source: @wakeupmbrc1, 2026-10-09
    • Without a signed VBIOS nobody sees how it would be possible. Source: @globalapathy, 2026-10-11
    • In theory 96 GB on some cards and 80 GB on most. Source: @djpandaman, 2026-10-12
    • The maximum confirmed on a production 8 GB card is 64 GB. Source: @djpandaman, 2026-10-12

Limits

  • No more SMs, no RIR path. Every RIR entry except the last reads 0xFFFFFFFF, so there is no software path past 74 SMs and 64 GB without NVIDIA-signed falcon code. Source: @wakeupmbrc1, 2026-10-07 Most cards have the full 128 SMs physically present. Source: @replacedriver47, 2026-10-07 Even with HS and debug mode, floorsweep values cannot change because of the DISABLE_SW_OVERRIDE fuse. Source: @cute_sc2, 2026-10-13 The 74 to 80 step that works on the 10 GB card does not work here: the extra TPCs sit behind OPT_TPC_DISABLE, which is burned to 1. Source: @djpandaman, 2026-10-10 The 8 GB card is "completely unfeasible on RIR paths". Source: @ob88594, 2026-10-27
  • Is anything more possible? Contested In the order posted:
    • Mountains keeps hitting locks that look wired into the silicon. Source: @mountains1605, 2026-10-20
    • matstatman thinks everything software-unlockable has been found and the fuses cut off the rest. Source: @matstatman, 2026-10-27
    • replacedriver47: fuses are read and propagated by software, so it is unlockable in theory if the security mechanisms can be bypassed. Source: @replacedriver47, 2026-10-27
  • PCIe Gen3. Blocked. The fuse register is locked, which blocks the Gen3 retrain; the RIR tool is not for 8 GB. Source: @wakeupmbrc1, 2026-10-25 OPT_DISABLE_GEN3_SPEED reads 1 on every card read so far, and forcing 8 GT/s on such a card drops it off the bus until a full power cycle. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20
    • Contested Can the last RIR record give single-fuse Gen3? mrpandagod could not enable Gen3 without several fuses and will not spend his last record to test. Source: @djpandaman, 2026-10-10 Rickcal believes only some cards can do single-fuse Gen3. Source: @rrickk420, 2026-10-10
    • A claim of Gen3 on an 8 GB card, with a release promised for 2026-12-09, came with no proof. Source: @rrickk420, 2026-10-09
  • PCIe width. The card links at x4 until the missing capacitors are fitted. Any x16 link, at any generation, needs the cap mod; x4 is the maximum without it. Source: @djpandaman, 2026-10-17 The mod is "borderline mandatory" for TP=2. Source: @cibernox, 2026-10-13 See PCIe capacitor mod.
  • NVLink (NVLink). No path on the 8 GB card. Source: @n2dcrypt, 2026-10-06 The same NVLink and power components are missing on all boards, Source: @globalapathy, 2026-10-24 and 8 GB NVLink has not been unlocked. Source: @ob88594, 2026-10-24
  • Open research (none of it usable yet): reading HBM site A's IEEE 1500 DEVICE_ID through the SEC2 Booter path, with disabled sites B and E next; Source: @junsulee119, 2026-10-24 using the last RIR record to change the device ID to A800, which would not change PCIe speed. Source: @djpandaman, 2026-10-10
  • Warning Unreleased "high power mode". matstatman found a register that enables a high power mode, also on 8 GB cards, does not know what it does and will not release it. Do not chase full-power modes on cards whose defective-marked stacks may run hot. Source: @matstatman, 2026-10-20

Best recipes

Full list with numbers and filters: All recipes. Engines and how recipes are written: Recipes. Recipe pages for this card include GLM 5.3 Flash on 4x 8 GB, DeepSeek V4.1 Flash on 4x 8 GB (Morrowmake), DeepSeek V4.1 Flash on 4x 8 GB (GNmac), GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 TP2) and GLM 5.3 Flash on 3x 8 GB. Numbers below are as posted; check the recipe page for engine, quant, context and concurrency.

  • 1 card (64 GB). Qwen3.8-27B W4A16 from a published Reddit recipe: about 170 tok/s decode reproduced; Source: @kellertreppe123, 2026-10-05 another owner reached about 3.7k tok/s prefill at about 100 tok/s decode. Source: @chonsworth, 2026-10-05 The current go-to single-card model is Qwen 3.8 Flash Next with PLE patches; the recommended recipe is "Qwen3.8-Flash-Next 3bpw MTP on 1x CMP 170HX SSD PLE AIO" (iIIusi0n). Source: @mountains1605, 2026-10-27
  • 2 cards (128 GB). Qwen 3.8 Flash Next: about 100-110 tok/s on 2 cards. Source: @68x0, 2026-10-09 A longer write-up by @plean. (Plean) exists; see Discord. GLM-5.3-Flash EXL3 in TP2 on two x16-modded cards: a longer write-up by @nestaa51 (Nestaa51) exists; see Discord.
  • 3 cards (192 GB). Qwen 3.8 Flash Next in PP=3 on unmodded x4 cards: a longer write-up by @c133 (clee) exists; see Discord. GLM-5.3-Flash W4A16 (about 4.3 bpw, 163 GiB) runs in PP3 with Morrowmake's recipe and 8-bit KV. Source: @shirogeek, 2026-10-19 An EXL3 GLM build scaled to 3 cards reached 60 tok/s in TP with full context, but the third card mostly buys context. Source: @wingrimez_40467, 2026-10-20
  • 4 cards (256 GB). Morrowmake's GLM 5.3 Flash recipe is the community daily driver: about 800 tok/s aggregate across agents, Source: @cibernox, 2026-10-19 130-160 tok/s single stream, Source: @mountains1605, 2026-10-20 317 tok/s at bs=1 on code and 164 on prose in TP4. Source: @turbine, 2026-10-27 For DeepSeek V4.1 Flash 1017, Morrowmake's recipe is rated above GLM Flash by Mountains after two days of agent work; Source: @mountains1605, 2026-10-26 greencandle's build streams the 189 GiB n-gram tables from NVMe. Source: @greencandle9753, 2026-10-21 The top DeepSeek decode numbers need the x16 mod. Source: @greencandle9753, 2026-10-22
  • Rule of thumb. Moving one card from x4 to x16 gave about +3% decode but about 4x prefill. Source: @wakeupmbrc1, 2026-10-07 With P2P working, PP still beat TP for very long-context prefill on one 4-card rig. Source: @mountains1605, 2026-10-15

Power, cooling and tuning

Details: Cooling and power.

  • Power input. One EPS 8-pin (CPU 8-pin), not PCIe. If you use the bundled 2x PCIe to EPS adapter, both PCIe inputs carry current. Source: @civiloid, 2026-10-15
  • Warning One cable into both adapter inputs. Feeding both inputs of the EPS adapter from one PCIe cable and its pigtail has burned hardware. Run a separate PSU cable to each adapter input, or use a CPU/EPS cable straight into the card. Source: @c133, 2026-10-19
  • Airflow. The card is passive and needs forced air even at idle; it falls off the bus at 95 C. Source: @analyticeth, 2026-10-07
  • Power cap. Most owners run 140-200 W. On GLM-5.3-Flash TP4 one club measured 412.7 tok/s single user at 140 W versus 418.6 at 165 W, with the best tokens per watt at 110 W; a driver reload resets the cap to the VBIOS default of 250 W, so reapply it at boot. Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20
  • Undervolt. Shifting the V/F curve cuts power by up to 30-35% at the same clock; it is per card, and instability shows up as Xid 31. Source: @wakeupmbrc1, 2026-10-12 One water-cooled 4-card setup runs 170tune at --ndiv 66 --offset 270 --clk 1410 --timings "REFRESH 24" at about 202 W. Source: @rh3dai, 2026-10-08 RH-3D also keeps a tuning video series for this card: a longer write-up by @rh3dai (RH-3D) exists; see Discord.
  • Idle. With the idle p-state patches, three cards idle at 23, 25 and 26 W. Source: @robbbbbb7973, 2026-10-10
  • 300 W. The 300 W VBIOS (92.00.6D.00.0A) exists for the 8 GB card only. Source: @zagitta, 2026-10-12

Known issues

Details: Troubleshooting, Driver and unlock failures, Instability under load.

  • 170tune preflight fails with "FBPA privilege masks are closed". One card out of 36 unlocked fine but refused 170tune with every unlocker and 170tune fork tried; the owner runs it on plain v0.5 without tuning. Source: @rh3dai, 2026-10-07 Another owner sees the same failure on all his cards and suspects a version mismatch between the unlocker and 170tune. Source: @drew_2, 2026-10-12
  • GSP init fails on every boot after Xid 119 and Xid 154. Reported on a card in a USB4 eGPU dock, persisting across hosts and power-off; unresolved. Source: @mistyashhole, 2026-10-10
  • Hangs under compute in a passthrough VM. A longer write-up by @vipic (ViPiC) exists; see Discord.
  • Whole rig loses power under heavy prefill. Four cards on one PLX board with two 750 W PSUs all dropped at once under big prompt-processing jobs; Source: @replacedriver47, 2026-10-05 nine cards on three 750 W PSUs at 160 W showed the same about once a day. Source: @reala7096, 2026-10-05
  • Warning Cards that fail to unlock. The reported overall failure rate for bought cards is about 20%. Source: @civiloid, 2026-10-13 Known failure modes: HBM memory errors, GSP that cannot initialise, and engineering-sample boards. Source: @ob88594, 2026-10-13 Cheap "8 GB" listings in China next to "64 GB guaranteed" ones are often cards with serious memory failures. Source: @replacedriver47, 2026-10-13

Buying

Details: Buying.

  • Test before paying: if the card is local, bring a bootable Linux USB and run a test unlock first. Source: @dottoreargus, 2026-10-11 A seller who refuses to test has a reason. Source: @ob88594, 2026-10-12
  • Check the device ID, VBIOS and board number for the 32 GB engineering-sample boards: Which card do I have#Cards that are not a normal 170HX.
  • Reported prices (USD unless noted, per card): about 550 for one owner's first four cards, bought before the October rush; Source: @gnasherrr, 2026-10-14 an eBay sale at 1,900 on 2026-10-01; Source: @djpandaman, 2026-10-12 3,000-3,500 by 2026-10-13; Source: @lenxtroz, 2026-10-13 in China ¥18,500 regular and ¥24,000 in bulk lots on 2026-10-22; Source: @ob88594, 2026-10-22 3,300-3,600 delivered with shipping and taxes from Chinese dealers on 2026-10-25. Source: @kellertreppe123, 2026-10-25

See also

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

Imported from the cmpunlocker Discord with credit; see Credits.
Retrieved from "https://cmp.wiki/w/index.php?title=8_GB_card&oldid=145"