Credits – CMP.WIKI

cmp.wiki is built from what the cmpunlocker Discord has worked out, and every fact on it carries the name of whoever posted it. Amogh, the server owner, cleared the message history for this.

People

358 people are credited on 1195 imported facts, ordered by how many facts credit them. One fact can credit several people.

Maintainers and main contributors

@_j3ns (Amogh): cmpunlocker and the cmpunlocker Discord server. 40 facts
@voidgiraffe (Morrowmake): the vLLM fork (vllm-cmp170hx) and the recipes built on it. 59 facts
@matstatman: the tools (gpcprobe, hbmmon). 106 facts
@wakeupmbrc1 (Miséricorde): Recipes, cooling and power, unlocks and drivers. HBM overclock via NDIV: 70 typical ceiling, 74 fails, gain depends on workload; Engine choice by card count: GLM-5.3-Flash W4A16 TP4 on 4 cards; Air cooling results: fan models, duty cycles and measured temperatures. 197 facts
@darkmaniac7 (Ivan): Recipes, unlocks and drivers, troubleshooting. Power caps reset to 250 W on every nvidia module load; PCIe Gen2 retrain works from inside the ESXi guest; ECC hardware recipe takes on 580: 14 protected writes. 175 facts
@djpandaman (mrpandagod): Unlocks and drivers, buying, card variants. 8GB card: 64GB is the confirmed ceiling, 80/96GB only theoretical; 74 SM is the hard ceiling on 8 GB cards; Capacitor mod soldering procedure. 156 facts
@globalapathy: Cooling and power, tools, card variants. PCIe x16 cap mod BOM: Samsung CL05B224KO5NNNC 0402 capacitors; Canonical cmpunlocker repo: amoghmunikote/cmpunlocker; BIOS prerequisites for 64 GB BAR1: Above 4G, ReBAR, a large MMIO-high window. 106 facts
@endsjeremy (Jeremy (Ends)): Platforms, cooling and power, buying. 64 GB DDR3 LRDIMMs only train on real LGA2011 server boards, not X99; Connector power ratings: PCIe 8-pin 150W spec, EPS 8-pin 300W continuous; Budget X99/Huananzhi route: DDR3 cuts RAM cost 4-5x. 102 facts
@mountains1605 (Mountains): Recipes, tools, community. Morrowmake GLM 5.3 Flash on 4x CMP 170HX PP4; Morrowmake/cmpunlocker adds working PCIe P2P; localairecipes.com: 170HX recipe leaderboard, filters. 78 facts
@cibernox: Recipes, cooling and power, unlocks and drivers. Prefill on CMP: TP4 is the weak layout; Morrowmake GLM 5.3 Flash recipe on 4x CMP; Undervolt / V-F offset stability: +150 stable, +250 gives garbage. 64 facts
@c133 (clee): Recipes, cooling and power, tools. clee’s Qwen3.8-Flash-Next recipe for 3x unmodded CMP 170HX, PP=3; clee recipe update 2026-10-11: persistent KV cache on NVMe across vLLM restarts; WARNING: ‘CMP-170HX-64GB-Unlocked-VBIOS’ repos ship a VBIOS identical to stock. 61 facts
@kellertreppe123: Buying, cooling and power, recipes. Qwen3.8-27B W4A16 on one 64GB 170HX: 160+/70+ tok/s decode; 10GB cards vanished from the market by 2026-10-07; Dell T5810 as a CMP host: special power cables, locked PWM on Linux. 56 facts
@sojun80 (Sojun): Recipes, buying, cooling and power. PRL mining figures on 170HX; Mining and serving share one card; DFlash speculative decoding on Qwen3.8-Flash-Next. 55 facts
@ob88594: Unlocks and drivers, buying, recipes. 8GB vs 10GB tradeoff: 10GB is now ‘A100 40G PCIe with more SM’; 10GB status: fully unlocked via the RIR method; PLX switches: MMIO issues force mailbox P2P. 53 facts
@gnasherrr: Cooling and power, platforms, recipes. DeepSeek V4.1 Flash on 4 to 8 CMPs: known recipes and their critique; CMP 170HX prefers int8/int4 quants; Water blocks: generic Ampere A100 PCIe blocks fit the 170HX, not 3090 blocks. 52 facts
@68×0 (68x): Unlocks and drivers, recipes, tools. ropuls gist: 2x CMP 170HX unlock plus cross-root P2P; Current go-to model for a single 64 GB CMP: Qwen 3.8 Flash Next with PLE patches; Flun GLM-5.3-Flash EXL3 3.05bpw repo scaled to 3 cards. 51 facts
@con3po7016 (Con3po): Cooling and power, recipes, buying. 64 GB is the 8 GB memory ceiling; Pearl mining on 170HX: 145-164 TH/s at ~200 W, economics and pools; GLM 5.3 Flash on 4 cards stable at 1.3M context. 48 facts
@lia_oc (_L_): Buying, unlocks and drivers, recipes. cmpunlocker v0.5 status: +4 SMs for 8GB and 10GB, ECC, no NVLink; Morrowmake dsv41-flash-cmp170hx-recipe: current best DS V4.1 Flash on 4x 8GB; 4-card recommendation: GLM 5.3 Flash w4a16 with the Morrowmake recipe. 46 facts
@reala7096 (reala): Recipes, cooling and power, troubleshooting. reala’s vllm-170hx-glm5 fork: full GLM 5.3 at 45 tok/s short; Unlock state only changes on a full cold power-off, not reboot, FLR or rescan; Bad card produces all-zero MoE rows at stock clock. 45 facts
@segmond: Recipes, buying, platforms. Octominer builds: 12-card X12 Ultra with external fans, and a 1x120mm plenum STL; x16 mod, TP and P2P are required to reach the published recipe numbers; DeepInfer and the HalfVulpes Deepinfer-SM80 port for CMP. 45 facts
@spectakkular (Takka): Buying, cooling and power, recipes. Memory leak in a prefill step of Morrowmake’s GLM 5.3 Flash runtime; 170HX vs A100 PCB and heat spreader differences that affect water block fit; 8GB supply estimate and the ‘binned defective memory’ argument. 39 facts
@alanoo: Cooling and power, recipes, unlocks and drivers. Engram ‘host’ mode can fail silently on DeepSeek V4.1 Flash; ECC counters under the driver patch; TensorFold added Ampere sm_80 CUDA kernels. 37 facts
@froggao (Froggo): Cooling and power, buying, platforms. Memory-error cards: HBM cannot be reworked; 8GB full SM unlock outlook: RIR path considered infeasible; GLM 5.3 Flash on 2 CMPs: said to run via Project Maya or any of the FLUN forks. 36 facts
@lxtend (lxwx): Unlocks and drivers, card variants, troubleshooting. lxwx confirms 0x00150095 clears DISABLE_SW_OVERRIDE; Xid 109 CTX SWITCH TIMEOUT at 128 SM with ‘defective’ GPCs; lxtend’s PCIe Gen3 unlock write-up and 170hx-rir-tool, 10GB only. 36 facts
@zingy90210 (PJ): Buying, platforms, recipes. Any x16 link needs the cap mod; Warning: never feed the 2×8-pin-to-EPS adapter from one PSU cable or its pigtail; vLLM KV cache offload to SSD for long-context reuse: clee’s 3x170hx config. 36 facts
@_orrorin (orrorin): Unlocks and drivers, recipes, cooling and power. RIR record encoding: bits [15:2] address, bit 1 data, bit 0 enable; RIR burn works only on 10GB cards: 8GB RIR rows are already blown at the factory; Strata tested on CMP 170HX: slower than vLLM, CMP not officially supported. 35 facts
@programmerboi: Cooling and power, platforms, recipes. SM counts: 170HX exposes 70/128 SM; Warning: stock heatsink needs forced airflow even at idle; Quant choice on Ampere/170HX: W4A16 via sm80 Marlin, INT4/INT8 supported. 34 facts
@_arcanum_ (Arcanum): Buying, cooling and power, card variants. Video: soldering the CMP 170HX x16 caps without blowing them away; Powering the 8-pin EPS input from server PSU breakouts; A100 40G/80G and 170HX share a PCB. 33 facts
@cute_sc2 (CUTE): Unlocks and drivers, buying, community. Qwen3.8 27B dense on one CMP 170HX: 210+ tok/s at 16k context; ECC reported available on unlocked CMPs; Saving a card with HBM errors: mark bad physical blocks at boot. 32 facts
@dottoreargus: Recipes, unlocks and drivers, cooling and power. dottoreargus prefill tuning on iIIusi0n vLLM: 5K+ tok/s on a single 8GB card; Adding CMPs to a consumer board: M.2-to-PCIe x16 extender cables; Untested: Noctua high-efficiency 120 mm fan grill for noise. 32 facts
@xeax: Cooling and power, tools, unlocks and drivers. Current driver for the unlock: nvidia-open 615.71.09; 10GB card after RIR burn: all non-defective SMs plus Gen3, memory 48GB at best; 170hx-rir-tool for 10GB PCIe gen / SM unharvesting. 32 facts
@rh3dai (RH-3D): Cooling and power, tools, unlocks and drivers. cmpunlocker v0.5 on 8GB: 64 GB, +4 SMs, ECC; cmpunlocker forks and 170tune lineages referenced in #testing-general; 170tune preflight FAIL on an 8GB card: FBPA privilege masks fenced. 30 facts
@ghostinthedatacenter_87281 (GhostInTheDatacenter): Card variants, unlocks and drivers, cooling and power. 10GB cards reach 128 SMs after the RIR burn; gpcprobe readout on 10 GB cards: GPC 5 healthy / 3 disabled / 0-1 defective; Idle draw: 33-40 W stock, 23-26 W with 170tune’s idle gate. 29 facts
@civiloid: Cooling and power, card variants, tools. Power input is one EPS 8-pin, not PCIe 8-pin; Used cards are filthy inside: plan to clean, repad and repaste when opened; Microsoldering tool kit used for the x16 cap mod. 26 facts
@zagitta: Unlocks and drivers, community, tools. cmpunlocker fails on nvidia-open 615.78.08; 300 W VBIOS exists for 8GB only; GA102 fuse and floorsweep register dump from two RTX 3090s. 26 facts
@phoenixone7 (Phoenix17): Recipes, buying, unlocks and drivers. Upgrading cmpunlocker to 0.5: rerun ./install.sh as root, no uninstall needed; Video generation on one CMP: 30s at 640×960 with H3; Model watch: DS V4.1 Flash 1017, Qwen 4, GLM 5.5 Flash, export-control rumor. 25 facts
@_syn_thetic_ (hawtt_dawgg): Cooling and power, buying, card variants. Stock paste is rock hard and pads are often missing; Bykski, TinyWhale and Alphacool 170HX water blocks leave bottom chokes; Cleaning filthy ex-mining cards: isopropyl rinses. 24 facts
@impactic: Recipes, cooling and power, unlocks and drivers. Eight cards: run two 4-card instances; Full GLM 5.3 tuned to 50 tok/s prose, 60 tok/s code on CMP 170HX; GLM Flash on 8 cards: 1600 tok/s, 2000+ with tuning and 230-250 W PL. 24 facts
@analyticeth: Cooling and power, troubleshooting, tools. Workaround for a stuck-at-zero HBM region: pin a 2 MB dummy allocation over it; Saving a card with HBM errors: mark bad physical blocks at boot; Never run a 170HX without a fan: card falls off the bus at 95C. 23 facts
@feralzi: Buying, unlocks and drivers, tools. Feralzi recipe: Qwen 3.8 27B at 225 t/s with 132k context on one CMP 170HX; SM unlock stuck at 80 of 126; RIR budget: 8 entries. 23 facts
@lycorascence: Card variants, unlocks and drivers, troubleshooting. 170hx-rir-tool prerequisites: cmpunlocker-patched driver, iomem=relaxed, root; Why 8GB unlocks lag: RIR records blown; State of the unlock: what each variant gets today. 23 facts
@markshowa: Cooling and power, platforms, tools. Supermicro 4028GR-TRT2 / 4028GR-TRT; P2P latency as a check: ~14 us means NCCL fell off P2P; Watercooling install guide in 170th-Street repo lacks pad and paste details. 23 facts
@rrickk420 (Rickcal): Unlocks and drivers, cooling and power, buying. Water block quality matters little at 175 W; Reference material: Consensus wiki, club-170hx, fuse RE notes, VBIOS tables; Warning: beware of fuse-burning ‘unlocks’ that can brick cards. 23 facts
@scallion_9883 (Humvee55): Unlocks and drivers, card variants, tools. Humvee55: 20x 10GB at 128 SM and Gen3 x16 with one RIR burn; Humvee55’s RIR/IFF notes; Unlock status Oct 2026: 10GB is at endgame, 8GB still blocked by poisoned RIR. 23 facts
@stcord (ST): Recipes, community, platforms. Engram/ngram placement is latency bound: RAM speed irrelevant; Tom’s Hardware covers the consumer-card port of the cmpunlocker compute unlock; Caution on deeper unlocks until permanent effects are ruled out. 23 facts
@lenxtroz: Buying, unlocks and drivers, recipes. Thermal pad thicknesses: VRAM 2 mm, MOSFETs 2 mm, coils 1 mm; Mixing a 170HX with other NVIDIA GPUs: Turing or newer only; cmpunlocker v0.5 on 8GB: 64 GB, +4 SMs, ECC. 22 facts

Everyone else on the Discord

@mlx4: Cooling and power, card variants, platforms. 2U cooling one 170HX at 250 W: 3 A 80 mm fans plus taped brackets took 10 C off; Rumor: 40×0 series unlocked like 30×0 for faster BF math; 8 GB vs 10 GB after unlock: 64 GB and 74 SM versus 40/48 GB, 128 SM and NVLink. 19 facts
@zorg33: Cooling and power, unlocks and drivers, card variants. Airflow data points and the S12038-4K push-pull consensus; Thermal limits: throttle from ~76 C, 85 C max, off the bus at 95 C; 250 W cap needs 2 to 3x the airflow of 200 W. 19 facts
@gpmickey (Mickey): Cooling and power, community, buying. 12-card room: dedicated breakers plus a focused air conditioner; 48 GB 4090 clamshell mods: watch backside VRAM temps; PCIe x16 cap mod BOM: Samsung CL05B224KO5NNNC 0402 capacitors. 18 facts
@n2dcrypt: Unlocks and drivers, card variants, cooling and power. Transient spikes reach 380 W at a 250 W cap; NVLink topology on A100-class PCIe cards: 3 bridge slots, 600 GB/s per pair; Open question: will NVLink be unlocked on the 8 GB card. 18 facts
@neapopa (Nea Popa): Cooling and power, platforms, buying. Hot-air cap mod technique: preheat, no added solder, tweezers hold; Warning: 6-pin to 8-pin adapter melted at 170W within weeks; Gigabyte G292-Z20 2U single-socket server for 8 CMPs: switch pairs, loud fans. 18 facts
@petrikrohn_26497 (Petri Krohn): Unlocks and drivers, platforms, card variants. Petri Krohn: unlocker does not need the debug booter_load; X99 DDR3 platform details: DDR3-capable v4 Xeons, board revisions; Why driver-loaded VBIOS overrides work on GA100. 18 facts
@replacedriver47: Unlocks and drivers, platforms, cooling and power. Whole rig loses power under heavy prompt processing; Minisforum MS-01 cannot allocate BAR1 for 4 cards; BAR1 stuck at 64 MB and P2P fails on Minisforum MS-01. 18 facts
@1a2b3c0967: Recipes, platforms, cooling and power. Morrowmake GLM-5.3-Flash recipe v1.7 user report: concurrency 3, 1M context; 8-card GPU servers: Gigabyte G292-Z20 and G481-HA0 give every card x16; Unlock work reported in progress: re-enabling ECC, PCIe Gen3, unlocking SMs. 17 facts
@programmer_007 (Programmer007): Recipes, cooling and power, buying. 85C is the thermal ceiling: diagnose with nvidia-smi, cap power, fix airflow; Fan power and PWM control without spare motherboard headers; vLLM on two 170HX: TP=2 usually beats one card. 17 facts
@shirogeek: Recipes, tools, unlocks and drivers. 8GB full-unlock paths: sw_fuse_override_disable register, dev-mode glitching; Benchmark hygiene: temp=0 and speculative decoding inflate shared tok/s numbers; GLM-5.3-Flash W4A16 on 3x CMP 64GB with Morrowmake’s recipe in PP3, 8-bit KV. 17 facts
@exile_1ck (eXile): Unlocks and drivers, card variants, recipes. Strata on a single 40GB 170HX: x4 to x16 mainly improves prefill; Strata engine: GPU-as-cache MoE offload; eXile’s custom vLLM build: Q3 Qwen 3.8 Flash Next on one 64GB 170HX. 16 facts
@nathang6866 (ElectroHobo): Recipes, platforms, tools. ElectroHobo Qwen3.8 Flash Next W4A16 recipe on 4x CMP with inco.ai drafter; ElectroHobo Qwen 3.8 Flash recipes for 2x and 1x CMP 170HX; Landscape of 2-card Qwen3.8-Flash-Next recipes on CMP 170HX: engines. 16 facts
@quark2world: Buying, recipes, platforms. Image and video generation on 170HX: Minimax H3 music video; Platforms for 10 cards at Gen4: PLX 88096 10-port switch on AM4, ROMED8; NVLink topology on A100-class PCIe cards: 3 bridge slots, 600 GB/s per pair. 16 facts
@_____josh_____ (Josh): Buying, card variants, community. One card in nine silently corrupted MoE rows at stock clocks; 10GB cards vanished from the market by 2026-10-07; 8GB supply estimate and the ‘binned defective memory’ argument. 15 facts
@cmdrwaka324 (Waka324): Platforms, troubleshooting, unlocks and drivers. Morrowmake GLM 5.3 Flash 1.7.x: looping symbols / Chinese text mid-reasoning; Morrowmake GLM5.3-Flash 1.6.0: one request’s prefill stalls another’s decode; Abliterated overlay for DS V4.1 Flash 1017 to avoid RE/binary-patching refusals. 15 facts
@ikyo: Recipes, cooling and power, platforms. GLM-5.3-Flash measured throughput on 4 cards; llama.cpp Q8 on a CMP lost to a Mac Studio on Qwen 27B; Single-card Qwen 27B under vLLM at concurrency 4. 15 facts
@squishysquare (You Know Who): Cooling and power, unlocks and drivers, card variants. In progress on 10 GB: 6th HBM stack enable and a 300 W BIOS; Debunked: screenshot claiming 192 GB / 130104 MB VRAM on a 10G card; Airflow data points and the S12038-4K push-pull consensus. 15 facts
@xanthefin (xanthe_): Card variants, unlocks and drivers, cooling and power. 300W power-limit bytes sit outside the VBIOS signature section; Warning: the 300W BIOS has already burned a 170HX; 300 W VBIOS 92.00.6D cross-flash: HBM 1728 MHz, burned-card report. 15 facts
@bigfoot1144 (bigfoot): Card variants, buying, recipes. FBP defect pattern is SKU-wide: every 10GB shows exactly 2 defective FBPs; GPC lottery: 8GB cards all have at least one defective GPC; Cap mod result stuck at x4, then x8 after cleaning caps. 14 facts
@nestaa51: Recipes, tools, troubleshooting. Nestaa51 recipe: GLM-5.3-Flash EXL3 TP2 with vision on 2x x16-modded 8GB CMPs; 2x 8GB cap-modded cards with P2P driver on Threadripper 128 GB; Reverse-engineering tool links shared: morluto/rea, fraktalcyber/lfi-rig. 14 facts
@port00 (RazK): Cooling and power, platforms, card variants. CMP 170HX integrated fan housing 3D model on Cults3D; Onboard fan header is vestigial: no PWM control, no fan curve in VBIOS; Hyve Edge Metal G10 EPYC server: x16 riser cables work in its PCIe 4.0 x24 slots. 14 facts
@shudza: Unlocks and drivers, platforms, recipes. Old dual-socket DDR3 workstations/servers as Strata hosts; 8GB vs 10GB tradeoff: 10GB is now ‘A100 40G PCIe with more SM’; vGPU close to working in cmpunlocker. 14 facts
@tekkenz: Card variants, recipes, unlocks and drivers. turboderp Qwen3.8-Flash-Next exl3 5.05bpw on a single 170HX; klee100 AutoRound 3bpw MTP Flash Next on iIIusi0n’s patched vLLM; Rameness: decisioning-based stored procedure harness. 14 facts
@vektor_prime (Ryko): Recipes, buying, community. vLLM run-to-run nondeterminism found while comparing Qwen3.8 Flash Next quants; vektorprime fork: ECC RAM changes merged; The ‘free 3090 speedup hack’ also works on RTX 3080 20GB and CMP 90HX. 14 facts
@boz._ (B-st): Troubleshooting, card variants, tools. cmp170hx-testsuite: open source test suite for incoming cards; Bad-HBM card passes basic Vulkan test but gives NaN inference; Reddit board-level repair of a dead 8 GB CMP. 13 facts
@dalzyyy (DaLZy): Buying, card variants, platforms. Even with the VBIOS signing key, 80/96GB is blocked by unknowns on the HBM side; FBP defect pattern is SKU-wide: every 10GB shows exactly 2 defective FBPs; Supermicro X11SPA-TF slot sharing and PLX topology. 13 facts
@fafinet: Platforms, cooling and power, troubleshooting. Fan test: two 40mm AVC DBTB0428B2G axials beat one Delta BFB1012VH 9733 blower; BMC cannot read GPU temperature of a CMP on a generic passive riser; Measured EPYC/TR host idle: ROMED8-2T + EPYC 7502 120 W at the wall. 13 facts
@greencandle9753 (greencandle): Recipes. DeepSeek V4.1 Flash 1017 on 4x 170HX: 150-296 tok/s single user; DeepSeek V4.1 Flash 1017 on 4 cards: GNmac-lab recipe versions and numbers; x16 hardware mod is the lever for top DSV4.1 decode numbers. 13 facts
@trustinter (Trust Inter): Unlocks and drivers, cooling and power, recipes. P2P fix: curiouspixel’s patch in Morrowmake/cmpunlocker; BAR1 cannot be assigned on ASUS WRX90; ASUS WRX80E/WRX90 boards: BAR1 stuck at 64 MB. 13 facts
@artistosu (gpulover): Buying, tools, community. DeepSeek V4.1 Flash on 4 to 8 CMPs: known recipes and their critique; 10GB card after RIR burn: all non-defective SMs plus Gen3, memory 48GB at best; Morrowmake roadmap: Qwen 3.8 Flash Next for 2 GPUs, exl3 kernel work. 12 facts
@busulle: Buying, unlocks and drivers. Huawei Atlas 300I Duo 96G is not a viable CMP alternative; 10GB full GA100 unlock via RIR fuse masking: exists privately, not released. 12 facts
@anon33anon: Recipes, unlocks and drivers, troubleshooting. x16 mod troubleshooting: card negotiates only x8 after cap mod; EXL3 at 3-3.5 bpw is close to 4-bit quants in quality on medium and large models; ExLlamaV3 DeepSeek V4.1 Flash support with per-card expert placement. 11 facts
@arabus_74064 (arabus): Unlocks and drivers, tools, platforms. 10GB card screenshot showing PCIe x16 Gen4; HBM temperature limits: throttle at 85C, keep under 80C, HBM damage risk; 10GB unlock working on Windows. 11 facts
@cuddylac997: Unlocks and drivers, buying, card variants. 10 GB unlocked memory: most cards stable at 40 GB, some reach 48 GB; cuddylac997’s private 10GB unlock: 1,741 GB/s stock bandwidth; Ten CMP 170HX in one host. 11 facts
@domrockt2: Buying, platforms, cooling and power. 3D-printed fan shrouds and adapters; Video generation on 8 GB cards: MiniMax H3 8 s clip in 10 to 15 min; Cold-boot alternatives for fuse testing: PCIe hot swap unreliable. 11 facts
@emogop: Unlocks and drivers, cooling and power, tools. CMP 90HX for AI: cap mod for x16, mod driver, 300-320 W limit; CMP 90HX VRAM plate mod: metal plate instead of pads keeps GDDR6X under 100 C; CMP 90HX: hardware ray tracing unlocked. 11 facts
@mrmannyatx (MrManny): Recipes, tools, platforms. MrManny’s qwen38-flash-next-2x-cmp170hx repo: 897K tokens of KV cache; mannydelrio recipe repo: Qwen3.8 Flash Next on 2x CMP 170HX; Harness choice: GLM 5.3 looping is usually a harness problem. 11 facts
@ropuls: Unlocks and drivers, platforms, cooling and power. P2P on CMP 170HX: cmpunlocker 0011 patch sets P2P override too early; P2P helps or hurts depends on topology: TP2 prefill doubles on EPYC; Tower cases for 5+ cards: Phanteks Enthoo Pro II Server Edition. 11 facts
@daveedcoles (Dc): Cooling and power, buying, platforms. Distros: unlock works on Proxmox; EPS12V CPU cables power the card; Bykski N-TESLA-A100-40G-SR block fits the newer CMP170HX. 10 facts
@peter643: Recipes, platforms, tools. peter643’s PEX88096 switch build; PEX88096 per-slot bifurcation via custom SBR firmware; vLLM checkpoint and fast model swapping. 10 facts
@zombielurker (Zombie): Troubleshooting, unlocks and drivers, cooling and power. Saving a card with HBM errors: mark bad physical blocks at boot; Does the 0.5 software ECC actually correct errors? Mixed evidence; PSU sizing data points and recommendations for multi-card rigs. 10 facts
@bigptoughneigh (Big Ptoughneigh): Unlocks and drivers, troubleshooting, cooling and power. NVLink: works with a simple non-public hardware mod; Gaming on 170HX: ~20 fps on Linux vs 60 fps on Windows. 9 facts
@cmurrayis (B16_c): Buying, cooling and power, unlocks and drivers. Custom single-slot 170HX water block project discontinued; State of the unlock: what each variant gets today. 9 facts
@dufrelic: Recipes, tools. DUFRelic’s qwen38-flash-next-recipe repo; Prefill on CMP: TP4 is the weak layout. 9 facts
@robbbbbb7973: Cooling and power, platforms, tools. Idle p-state patches: 23-26 W idle, 170tune idle gate config; Idea: drop PCIe link speed at idle for extra idle savings. 9 facts
@turbine: Recipes, cooling and power, tools. turbine: Qwen3.8-27B TP4 on 4x 170HX at 440 tok/s decode, ~4k tok/s prefill; Prefill on CMP: TP4 is the weak layout. 9 facts
@ichundes (Ago): Troubleshooting, unlocks and drivers, tools. Undervolt / V-F offset stability: +150 stable, +250 gives garbage; Dead-on-arrival 10 GB card: VBIOS ROM reads 0xFFFF, GSP init fails. 8 facts
@jjb01957 (jjb): Recipes, cooling and power, buying. Five threat layers for local AI deployments; Morrowmake dsv41-flash-cmp170hx-recipe: current best DS V4.1 Flash on 4x 8GB. 8 facts
@waypoint3642 (Waypoint): Unlocks and drivers, troubleshooting. Waypoint: 10GB with all non-defective SMs enabled; Xid 109 CTX SWITCH TIMEOUT at 128 SM with ‘defective’ GPCs. 8 facts
@adusts (Adust): Buying, community. Gaudi 2 OAM-to-PCIe works electrically. 7 facts
@colo2442 (Colo): Card variants, unlocks and drivers, troubleshooting. 10G unlock: enable defective GPCs but mask only the faulty TPCs; 10G with more than about 80 SMs: self-reported power becomes unreliable. 7 facts
@danske1 (Danske): Cooling and power, buying, card variants. Arctic fan rewiring on the stock shroud: quieter, thermals unknown; Hot-air cap mod technique: preheat, no added solder, tweezers hold. 7 facts
@legitsplit: Buying, platforms, cooling and power. Huananzhi H12-8D reported at about 50 W idle; Tensor un-nerf reported on Ada via gpu-unlock forks. 7 facts
@meatsus: Buying, platforms, unlocks and drivers. State of the unlock: what each variant gets today; Supermicro X11SPA-TF slot sharing and PLX topology. 7 facts
@moliminous: Cooling and power, troubleshooting, recipes. RazK’s integrated fan housing: stock-size air mod, 5 cards averaging 33-37 C; 5x 170HX on a GPU switch baseboard dropping off PCIe on large prompts. 7 facts
@vcore2191 (Vcore): Cooling and power, unlocks and drivers, tools. LACT curve editor accepts V/F curves 170tune refuses; Undervolt / V-F offset stability: +150 stable, +250 gives garbage. 7 facts
@armand0e: Unlocks and drivers, card variants, troubleshooting. 8 GB vs 10 GB: the only RIR difference is slot availability; 8GB Gen3 status: fuse register locked, blocks the gen3 retrain. 6 facts
@chonsworth: Troubleshooting, platforms, unlocks and drivers. ECC enable patch does not change usable VRAM; Dell R740: suspected dead GPU power plug on the system board. 6 facts
@drew_2 (drew-3): Unlocks and drivers, troubleshooting, tools. nvidia-open 615.71.09 is the working driver; Driver 615.78.08 breaks the unlock. 6 facts
@endbr64 (Zoodasa): Buying, card variants. 6 facts
@fiiireface: Cooling and power, recipes, community. Open question: GA100 die dimensions for cutting PTM7950; mannydelrio recipe repo: Qwen3.8 Flash Next on 2x CMP 170HX. 6 facts
@finishtilt14: Cooling and power, recipes, platforms. Warning: aged ZSX-AMPS breakout boards caused reboots and PSU shutdowns; Gigabyte MZ31-AR0: one physical x16 is electrically x8. 6 facts
@galygious (GalyGames): Platforms, recipes, buying. Uncensored variants in use: huihui abliteration on Qwen3.8 Flash Next; Plean recipe: Qwen3.8 Flash Next on 2x CMP 170HX with vLLM PP. 6 facts
@gitanos: Buying, community. Benchmark standardisation effort: llm-inference-bench, 1/2/4/8-card matrix. 6 facts
@junsulee119 (Jason): Unlocks and drivers, community, buying. 8GB: HBM site A IEEE 1500 DEVICE_ID read via SEC2 Booter write permission; 8 GB card beyond 64 GB: disabled-but-not-defective FBP masks. 6 facts
@sh.ad (shad): Unlocks and drivers, recipes, tools. 3D-printed shroud air cooling: seal the gaps; Current go-to model for a single 64 GB CMP: Qwen 3.8 Flash Next with PLE patches. 6 facts
@zarquon3157 (Zarquon): Unlocks and drivers, buying, cooling and power. mrpandagod: Gen2 was a bootstrap bypass; 8GB Gen3 via the single remaining RIR record: contested. 6 facts
@deejay9983 (DeeJay): Recipes, buying. 6x 170HX DeepSeek V4.1 Flash on x4 links: 67-90 tok/s decode; x16 hardware mod is the lever for top DSV4.1 decode numbers. 5 facts
@douya4305 (douya): Card variants, unlocks and drivers, buying. PG199-origin ‘170HX’ engineering-sample cards are 32 GB: subsystem 10de:1586; 8GB full-unlock rumors circulating Oct 10-14: unverified. 5 facts
@fnerdman: Platforms, cooling and power, buying. HBM overclock via NDIV: 70 typical ceiling, 74 fails, gain depends on workload; 300 W VBIOS 92.00.6D cross-flash: HBM 1728 MHz, burned-card report. 5 facts
@green_dam (GreenDam): Unlocks and drivers, card variants. NV_FUSE_OPT_SM_SPEED_SELECT_FMLA32 / IMLA4 read 0x5 on 90HX vs 0x1 on RTX 3090; P106-090 also runs PCIe Gen2 x16. 5 facts
@mfj4: Buying, cooling and power, tools. A30 unlockability speculation: same GA100 die, heavily harvested, 24 GB; FSP U500 case with a custom sliding plenum for 2 to 4 CMPs. 5 facts
@nikhilr1696 (NikhilR): Cooling and power, tools, unlocks and drivers. Never run a 170HX without a fan: card falls off the bus at 95C; Capacitor mod: what it is, who will do it, and what it costs. 5 facts
@reggie7756 (regaly): Card variants, recipes, buying. 64 GB is the 8 GB memory ceiling; Qwen3.8 27B dense on one CMP 170HX: 210+ tok/s at 16k context. 5 facts
@sekkit_75019 (aidoc): Unlocks and drivers, community, tools. gen3_sku_probe.py: 10GB card has all three Gen3 fuses set; aidoc working on Gen3 for the 8GB SKU. 5 facts
@tugoubxs: Recipes, buying, unlocks and drivers. 10GB status: fully unlocked via the RIR method; Quant choice on Ampere/170HX: W4A16 via sm80 Marlin, INT4/INT8 supported. 5 facts
@vipic: Troubleshooting, recipes, platforms. Debian 13 build fix: ‘stdarg.h: No such file or directory’; Install recipe: install.sh flags, prerequisites, verify steps. 5 facts
@xtj7: Tools, unlocks and drivers, cooling and power. Air cooling results: fan models, duty cycles and measured temperatures; Upgrading cmpunlocker to 0.5: rerun ./install.sh as root, no uninstall needed. 5 facts
@ybryany: Buying, troubleshooting, unlocks and drivers. CMP 170HX works beside an RTX 4080 Super on CachyOS; Memory-error cards: HBM cannot be reworked. 5 facts
@zefrench6961 (Zeferrum): Card variants, unlocks and drivers, tools. PG199-origin ‘170HX’ engineering-sample cards are 32 GB: subsystem 10de:1586; Spotting the 32GB engineering-sample card before buying. 5 facts
@.henrykrinkle (HenryKrinkle): Unlocks and drivers, cooling and power. ECC tool 0.5 fails on driver 615.78.08, works on 615.71.09; Warning: never feed the 2×8-pin-to-EPS adapter from one PSU cable or its pigtail. 4 facts
@atsasin2126 (atsasin): Card variants, platforms. HBM2 vs HBM2e on the 10GB, stack counts and bus width; The ’80GB on 10GB’ paper: the extra capacity is aliasing, not memory. 4 facts
@catchyuser (Kaatchi): Unlocks and drivers, platforms. ECC tool 0.5 fails on driver 615.78.08, works on 615.71.09; Supermicro GPU servers as hosts: SYS-4028GR and 1028GQ-TRT noise and idle draw. 4 facts
@d3m3n7r: Troubleshooting, tools, unlocks and drivers. Code 43 / ‘DQ memory training failed’ traced to a failed memory chip; Tool link: nvidia-pstated daemon for automatic P-state management. 4 facts
@dany0_ (Daniel): Buying, card variants, unlocks and drivers. 10GB cards vanished from the market by 2026-10-07; Outlook on remaining unlocks: NVLink hard, PCIe Gen3 ‘not consistent enough’. 4 facts
@giaki3003: Unlocks and drivers, buying, community. cmpunlocker Rev 0.5 released: software-based ECC on the master branch; Debunked: screenshot claiming 192 GB / 130104 MB VRAM on a 10G card. 4 facts
@intervitens: Recipes, buying, platforms. Card count and parallelism: vLLM TP must divide the attention head count; Unbinding the CMP from the driver for the RIR burn: sysfs unbind, blacklist. 4 facts
@landontg (Landon): Buying, cooling and power. Fan choice and noise: Arctic P12 Pro / S12038 4K-8K preferred over Noctua A12x25. 4 facts
@newman_08462 (Newman): Card variants, buying, tools. FBP defect pattern is SKU-wide: every 10GB shows exactly 2 defective FBPs; GPC lottery: 8GB cards all have at least one defective GPC. 4 facts
@okrasit: Troubleshooting, cooling and power, recipes. PSU sizing data points and recommendations for multi-card rigs; Morrowmake GLM 5.3 Flash 1.7.x: looping symbols / Chinese text mid-reasoning. 4 facts
@pcsv: Unlocks and drivers, cooling and power, recipes. lxtend’s PCIe Gen3 unlock write-up and 170hx-rir-tool, 10GB only; 8GB Gen3 status: fuse register locked, blocks the gen3 retrain. 4 facts
@plean: Recipes, tools. Plean recipe: Qwen3.8 Flash Next on 2x CMP 170HX with vLLM PP; vLLM run-to-run nondeterminism found while comparing Qwen3.8 Flash Next quants. 4 facts
@poningru: Unlocks and drivers. Unlocking more memory on a real A100 40GB: 48 GB via the 10GB-card method. 4 facts
@riko6253 (riko#riko): Unlocks and drivers. lxtend’s PCIe Gen3 unlock write-up and 170hx-rir-tool, 10GB only; RIR burn works only on 10GB cards: 8GB RIR rows are already blown at the factory. 4 facts
@seriousrising: Unlocks and drivers, tools. Full P2P via Morrowmake’s fork patch; Sidecar GPU for the speculative-decoding draft model. 4 facts
@skirtchaser: Cooling and power, card variants, buying. IHS thermal interface: PTM7950 vs Laird TPCM 7000 phase change material; Under the IHS is phase change paste, not solder. 4 facts
@tokukarin: Card variants, buying, recipes. Qwen Flash Next on one 8GB-variant card with llama.cpp. 4 facts
@a1rm4x: Cooling and power, platforms, community. In-progress builds and promised follow-ups; Gigabyte G292-Z20 2U GPU server as a CMP platform. 3 facts
@akkerstone (David): Buying. PRL mining figures on 170HX; Mining data point: unlocked 10GB 250 TH at 250W vs 8GB 165 TH at 1170 MHz core. 3 facts
@amishrakefightr: Recipes, buying. Qwen3.8-27B W4A16 on one 64GB 170HX: 160+/70+ tok/s decode; Qwen 3.8 27B single-card numbers by quant and engine: INT4 ~55. 3 facts
@backwardsminer (ᗺackwardsMiner): Cooling and power, platforms. 3D-printed fan shrouds and adapters; Large rigs: 16 cards per server is the practical ceiling. 3 facts
@bismuth67 (benoit): Platforms. VM passthrough: unlock on the Proxmox host, then pass through; Hypervisor passthrough: unlock in-guest is finicky. 3 facts
@chris0chris0chris: Card variants, buying. CMP 90HX for AI: cap mod for x16, mod driver, 300-320 W limit. 3 facts
@da6b (dagb): Card variants, tools, recipes. CMP 90HX has no unlockable memory; Card count and parallelism: vLLM TP must divide the attention head count. 3 facts
@humandrone8721_60165 (HumanDrone8721): Unlocks and drivers, buying. Capacitor mod soldering procedure; Capacitor mod: what it is, who will do it, and what it costs. 3 facts
@ju5t5omedude: Cooling and power. 3D-printed shroud air cooling: seal the gaps; RTX 3050 6GB gpu-unlock result: FP16/BF16 tensor rate doubled, GEMM check clean. 3 facts
@locs_locs (Kachowtowmater): Buying, card variants. 8GB vs 10GB: which variant to buy. 3 facts
@lukextra (Luke Howard): Platforms, card variants. Motherboard options for 6-card P2P: WRX80, ROMED8-2T/SP3, H12SSL-i; 8GB cards: stock Gen2 x4 link, no PCIe 4.0 benefit. 3 facts
@marklintonca (Mark Linton [CA]): Troubleshooting, card variants, unlocks and drivers. RIR macro on 8 GB cards is 15/16 used: every RIR; Warning: writing 0x15011401 to RIR row 0 does nothing and wastes a row. 3 facts
@mistyashhole: Troubleshooting, unlocks and drivers. 8GB card fails GSP init on every boot after Xid 119/154; cmpunlocker v0.5 on 8GB: 64 GB, +4 SMs, ECC. 3 facts
@mowgli8628 (dash): Troubleshooting, recipes, unlocks and drivers. Decode dropped from 60-80 to 35 t/s after moving from cmpunlocker 0.4 to 0.5; Qwen 3.8 Flash Next vs 27B for coding in 64GB. 3 facts
@persona_80337 (PersonA): Buying, recipes. First 4-card GLM-5.3-Flash run: 162 tok/s single user. 3 facts
@prefix10k (DOG): Cooling and power, card variants. Tesla V100 not a viable alternative: old, no current driver support; Cooling: AMD V620 80mm fan shrouds plus Arctic S8038-10K fans. 3 facts
@smd386: Recipes. Card count and parallelism: vLLM TP must divide the attention head count; DS V4.1 on 4 cards needs ~2.9bpw. 3 facts
@tchelobh: Tools, platforms, unlocks and drivers. 10GB cards reach 128 SMs after the RIR burn; iff_apply.py, 170hx_sweep.py and bar0_write.py drop, marked untested. 3 facts
@thanos8912 (Thanos): Community, buying, cooling and power. Thanos is building an inference marketplace served from 170HX cards; Power limit: unlocked 10G/full-SM cards power throttle. 3 facts
@youngju10000 (영쥬): Troubleshooting. Dead-on-arrival 10 GB card: VBIOS ROM reads 0xFFFF, GSP init fails; Warning: VBIOS TOCTOU test reportedly bricked a card permanently. 3 facts
@zaknafar: Cooling and power, buying. Bykski N-TESLA-A100-40G-SR block fits the newer CMP170HX. 3 facts
@aaasss1456 (aaasss): Buying, card variants. 8GB ’80/96 GB’ rumor: screenshots from Chinese sites and Twitter, unverified. 2 facts
@acedogblast: Recipes, unlocks and drivers. Qwen3.8 Flash Next on a single CMP 170HX with Strata: more than 60 tok/s. 2 facts
@bayley_w: Buying. 2 facts
@cannibaltoast: Platforms. Strata-style engines only help MoE models, not dense models. 2 facts
@catcat0458 (catcat): Tools. hbmmon monitoring tool shared. 2 facts
@cressida1881 (cressida): Cooling and power, card variants. Hot-air cap mod technique: preheat, no added solder, tweezers hold. 2 facts
@dale063629 (大乐): Tools, recipes. Recipe repo: DeepSeek V4.1 Flash on CMP 170HX by wgkbfcp. 2 facts
@dkpoulsen (thcleaner): Tools, card variants. 170hx-fanduct: OpenSCAD generator for multi-card fan ducts. 2 facts
@fdelfini (Fernando Delfini): Platforms. ASUS P9X79 WS: unlock needs Above 4G decoding and a larger MMIOH window. 2 facts
@forte9000: Cooling and power, buying. 2 facts
@galygious2: Card variants, recipes. Qwen 3.8 Flash Next on CMP: 4-card and 1-2 card recipes and numbers. 2 facts
@givisok: Cooling and power. San Ace S12038-8K run at 20 to 25 % speed with a 3D-printed finger grill. 2 facts
@golden_panda_97088_61593 (OW3): Buying. 2 facts
@grapepepper: Tools. Tensor un-nerf reported on Ada via gpu-unlock forks. 2 facts
@hezz1891: Unlocks and drivers. 170hx-rir-tool: Gen3 confirmed on 10GB, Gen4 payload untested, 10GB only. 2 facts
@iksboj (Ashrel): Unlocks and drivers. Open question: does the unlocker apply to the RTX 3080 20GB?. 2 facts
@ins_mniac (insomniac): Community, cooling and power. Proposed Chinese export controls on AI models: download weights now. 2 facts
@javicabrera_30495 (JCP): Tools, card variants. CMP 90HX: 3090-level compute when unlocked but 10GB VRAM, no NVLink, no RIR rows. 2 facts
@jihoon_yi (felix-e): Buying, tools. Strata supply-chain warning: it pulls in further external libraries. 2 facts
@jimtjames (gym): Buying, unlocks and drivers. int8 tensor-core throughput stays gated after the unlock. 2 facts
@jma1753 (JMA1976): Tools. CMP 100-210 Tensor-only unlock repo works for one user. 2 facts
@kawaii8789 (kawaii): Platforms, recipes. Pearl mining on 170HX: 145-164 TH/s at ~200 W, economics and pools. 2 facts
@killred (vks): Troubleshooting. Morrowmake GLM 5.3 Flash 1.7.x: looping symbols / Chinese text mid-reasoning. 2 facts
@larrylegend3504 (larrylegend): Cooling and power, card variants. 170HX vs A100 PCB and heat spreader differences that affect water block fit. 2 facts
@leenodie: Cooling and power. Water cooling: A100-style blocks fit. 2 facts
@licensedterrapin (LT): Recipes. Model picks: Breeze for one-shot voice cloning, Qwen 2.1 for image editing. 2 facts
@logy0 (Logy): Platforms. Proxmox/QEMU GPU passthrough guide used for 170HX. 2 facts
@lubakmanqk (Lyubo): Buying, recipes. 6x 170HX DeepSeek V4.1 Flash on x4 links: 67-90 tok/s decode. 2 facts
@manicraft1001: Buying. 2 facts
@monroestreet: Buying. Open questions raised in this chunk. 2 facts
@n3bs: Unlocks and drivers. Running gpu-unlock and cmpunlocker on the same machine. 2 facts
@odide_36591 (GPTGOLDTrader): Buying, community. 2 facts
@onemillionkates (kater): Unlocks and drivers, card variants. NVLink on 170HX: the board lacks U11, resistors and power phases. 2 facts
@pekoraali: Cooling and power. 3-card air-cooled ATX ‘sleeper’ build: P14 Pro push-pull, 120W TDP, 69-79C. 2 facts
@redpandamining: Platforms, unlocks and drivers. NVIDIA driver 615.71.09 confirmed working with cmpunlocker v0.5 on 10GB cards. 2 facts
@roboalchemist: Cooling and power, recipes. roboalchemist FP4 DeepSeek V4.1 1017 mix for 6 CMPs. 2 facts
@supasnipa360: Unlocks and drivers, tools. Warning: the ‘CMP-170HX-64GB-Unlocked-VBIOS’ GitHub repos are bogus. 2 facts
@tihohod_ (tihohod): Unlocks and drivers, recipes. Install recipe: install.sh flags, prerequisites, verify steps. 2 facts
@truededinfo (DEDinfo): Community, unlocks and drivers. 10GB SM unlock: 74 to 80 SMs without fuse override. 2 facts
@unlikelypotato: Platforms, cooling and power. Cooling: AMD V620 80mm fan shrouds plus Arctic S8038-10K fans. 2 facts
@wingrimez_40467 (Wingrime): Card variants, recipes. Flun GLM-5.3-Flash EXL3 3.05bpw repo scaled to 3 cards. 2 facts
@zarknoth: Unlocks and drivers. GSP disabled with a dual-stack driver keeps the unlock across driver reload. 2 facts
@_kanorin (Kanorin): Platforms. Gigabyte MZ32-AR0 runs up to 14x 8GB CMP 170HX. 1 fact
@aaron0010 (aaron): Recipes. Recipe repo: DeepSeek V4.1 Flash on CMP 170HX by wgkbfcp. 1 fact
@adrienbrault: Cooling and power. 3D-printable parametric 120 mm fan duct for 4 or 5 PCIe slots. 1 fact
@akuchen_66498 (AKuchen): Unlocks and drivers. RIR fuse bypass path: works in principle on 10GB cards. 1 fact
@aleksandartsvetanov: Recipes. Quant quality for Qwen3.8 Flash Next: prefer 4-bit G32 over G128. 1 fact
@aluminer_ (Aluminer): Unlocks and drivers. Upgrading cmpunlocker to 0.5: rerun ./install.sh as root, no uninstall needed. 1 fact
@bagstah (Bagster): Run coding agents in a sandbox. 1 fact
@bigpoppag91 (BigPoppaG): ESP32 ‘unlock’ tangent: microcontrollers, not AI accelerators. 1 fact
@bingo.fuel: CMP 90HX for AI: cap mod for x16, mod driver, 300-320 W limit. 1 fact
@bvls7857 (BVls): Community. 1 fact
@christopherbd (_CBD): Platforms. 170HX as an Oculink eGPU on a Dell 7040m micro PC: works, about 50 t/s. 1 fact
@curiouspixel: Unlocks and drivers. Full P2P via Morrowmake’s fork patch. 1 fact
@damnitcharlieai (Kachowtowmater): Card variants. PCIe x16 capacitor mod: needs microscope and hot-air station. 1 fact
@daniyel0042 (Daniyel): Unlocks and drivers. Mixing a 170HX with other GPUs and which driver versions work. 1 fact
@darkshad03 (DarkShad0w): Unlocks and drivers. gen3_sku_probe.py: 10GB card has all three Gen3 fuses set. 1 fact
@deaddudehangin: Unlocks and drivers. cuddylac997’s private 10GB unlock: 1,741 GB/s stock bandwidth. 1 fact
@deced_123 (deced): Troubleshooting. Xid 109 CTX SWITCH TIMEOUT at 128 SM with ‘defective’ GPCs. 1 fact
@evcheine (amongodb): Gaming on the 170HX: works poorly, not worth it. 1 fact
@f0ol (Andy): Recipes. 1 fact
@fallengod9962: Platforms. GLM 5.3 on 2 cards at PCIe x4 is brutally slow. 1 fact
@farawaythrower (Far): Cooling and power. PSU sizing data points and recommendations for multi-card rigs. 1 fact
@fedzel: Platforms. 1 fact
@golder555 (Leviathan of freedom): Community. 1 fact
@hamuki: Buying. 1 fact
@hogg7701 (hogg): Recipes. dottoreargus prefill tuning on iIIusi0n vLLM: 5K+ tok/s on a single 8GB card. 1 fact
@honzapat_cz: Gaming on the 170HX: works poorly, not worth it. 1 fact
@hum1d: Tools. Validation on arrival: club-170hx repo steps, gpcprobe. 1 fact
@info.at.mah7_78733 (MAH7): Tools. Agent computer-use over IP-KVM: Con3po’s setup docs and related projects. 1 fact
@jack0x01: Platforms. VM passthrough: unlock on the Proxmox host, then pass through. 1 fact
@jas0xf: NVLink on the 170HX: expected value only for diffusion. 1 fact
@jeffmcjunkin: Buying. 1 fact
@johnytonny (Johny Tonny): Unlocks and drivers. 10GB card can reach PCIe Gen4, 8GB stuck at Gen3. 1 fact
@jonathanyjoerrington (Jonathany Joerrington): Recipes. turboderp Qwen3.8-Flash-Next exl3 5.05bpw on a single 170HX. 1 fact
@justinjja: Recipes. DeepSeek V4.1 Flash on 4 to 8 CMPs: known recipes and their critique. 1 fact
@kamilake (Kamilake 카미): Buying. 1 fact
@kill3again (Q3): 3080 Ti 20G is a 320-bit 20GB variant of the 3090. 1 fact
@kio063218 (kio): Community. Debunked: screenshot claiming 192 GB / 130104 MB VRAM on a 10G card. 1 fact
@lamune6809 (lamune): Cooling and power. Open server slots: populate all slots or print shrouds vs just seal the openings. 1 fact
@m3gagluk: ESP32 ‘unlock’ tangent: microcontrollers, not AI accelerators. 1 fact
@mellojello_ (Mello Jello): Tools. Coding harnesses and clients people pair with local models: dsh, maki. 1 fact
@minioadmin: Card variants. CMP 90HX: 3090-level compute when unlocked but 10GB VRAM, no NVLink, no RIR rows. 1 fact
@mitakihara_ (Mitakihara): Unlocks and drivers. Unlock status Oct 2026: 10GB is at endgame, 8GB still blocked by poisoned RIR. 1 fact
@mrpec (Mr Pec): Unlocks and drivers. 1 fact
@nioroso: Tools. LMCache CPU KV offload with vLLM: unclear benefit, serialized requests. 1 fact
@nyancat3355 (NyanCat): Unlocks and drivers. Unlock status Oct 2026: 10GB is at endgame, 8GB still blocked by poisoned RIR. 1 fact
@paul_p9524: Cooling and power. WRX80 fan headers: 12 V 2 A each. 1 fact
@predictorandrzej (Predictor Andrzej): Tools. Identifying unlockable units: gpcprobe, reading the RIR, infinite.im SM table. 1 fact
@ratdoux: Buying. 1 fact
@ronnyfish: Unlocks and drivers. 8GB Gen3 via the single remaining RIR record: contested. 1 fact
@sec_dude (secdude): Community. 1 fact
@siriusrhi (Sirius (Rhi)): Unlocks and drivers. 8GB Gen3 status: fuse register locked, blocks the gen3 retrain. 1 fact
@solada_28598 (Solada): Buying. 1 fact
@spiralglide: Platforms. VM passthrough: unlock on the Proxmox host, then pass through. 1 fact
@stormlrd: Cooling and power. Thermal interface refresh candidates: PTM7950 or Laird A18172-08 phase change. 1 fact
@supermanisdum: Cooling and power. Repasting: IHS with paste underneath, big paste volume, typical gain ~2 C. 1 fact
@tarxlazer: Unlocks and drivers. 8GB full-unlock rumors circulating Oct 10-14: unverified. 1 fact
@tekfabii (YellowstoneCrypto): Community. Open question: unlocks for other CMP models. 1 fact
@truneski: Buying. 1 fact
@vuhuycan0322 (vuhuycan): Idea: host2jtag / scan-chain brute forcing of fuse overrides. 1 fact
@wise_axolotl_13963 (dly): Tools. Agent computer-use over IP-KVM: Con3po’s setup docs and related projects. 1 fact
@xiaomi1234520 (Crxpted): Troubleshooting. Dead-on-arrival 10 GB card: VBIOS ROM reads 0xFFFF, GSP init fails. 1 fact
@xnupanic76: CMP 90HX for AI: cap mod for x16, mod driver, 300-320 W limit. 1 fact
@youcancallmenicki (You can call me Nicki): Unlocks and drivers. 170hx-rir-tool: Gen3 confirmed on 10GB, Gen4 payload untested, 10GB only. 1 fact

Credited from GitHub and elsewhere

Consensus-Protocol on GitHub: Card variants, unlocks and drivers, platforms. Two SKUs: 8 GB to 64 GB and 10 GB to 40 GB; PCIe Gen3/Gen4 blocked by OTP fuses; What survives FLR vs power cycle. 20 facts
PixelML on GitHub: Recipes, platforms, unlocks and drivers. Proxmox/VFIO passthrough: use SeaBIOS, not OVMF; Failure signatures and recovery; Pipeline-parallel beats tensor-parallel on the no-NVLink fabric. 19 facts
bayley on GitHub: Platforms, unlocks and drivers, cooling and power. Mailbox P2P is a dead end; How cmpunlocker reaches PCIe Gen2: the registers, PLM gates and retrain sequence; Bare-metal static-BAR1 P2P recipes and measured bandwidth. 15 facts
lixiangwuxian on GitHub: Tools, unlocks and drivers, card variants. 170hx-rir-tool: what RIR is, what –write-rir burns; GA100 RIR fuse writer. 8 facts
admunch888 on GitHub: Recipes, unlocks and drivers, platforms. admunch888 fork: –mclk-ndiv, –refresh, p2p-verify, ECC opt-in; Mailbox P2P is a dead end. 7 facts
LingzheZhao on GitHub: Community, unlocks and drivers, recipes. Bare-metal static-BAR1 P2P recipes and measured bandwidth; cmpunlocker ECC is three patches. 6 facts
satspace-cpu on GitHub: Troubleshooting, recipes, unlocks and drivers. Verified static-BAR1 P2P bandwidth; MMA/Tensor-core unlock verification method. 6 facts
ahnguyen17 on GitHub: Card variants, tools, unlocks and drivers. 8 GB vs 10 GB: the only RIR difference is slot availability; Recoverable FBPs and the 80 GB FF-append theory. 5 facts
cebeuq on GitHub: Recipes, community. ECC ‘proven by multiple people’ rests on the reporting patch; Community HBM clock numbers: 1.59 vs 1.32 TB/s, +6% single user on a TP4 host. 5 facts
lesj0610 on GitHub: Troubleshooting, community, platforms. P2P failure modes on bare metal: silent corruption without trap-31; Fork ecosystem map: who carries what beyond upstream cmpunlocker. 5 facts
rjohny55 on GitHub: Unlocks and drivers, platforms, community. rjohny55 bounded Gen2 second-pass method; cmpunlocker ECC is three patches. 5 facts
Zanooda on GitHub: Recipes, troubleshooting. Zanooda DeepSeek V4.1 Flash PP8 recipe; Card falls off the bus at the stock 250 W cap. 5 facts
anoane on GitHub: Platforms, recipes. Hypervisor passthrough: unlock in-guest is finicky; DeepSeek V4.1 Flash recipe landscape for CMP 170HX and community numbers. 4 facts
cachenetics on GitHub: Cooling and power, troubleshooting. 170tune HBM overclock: two ceilings; 170tune undervolt: VF offset and clock ceiling. 4 facts
dev-boz on GitHub: Troubleshooting, tools. QC acceptance ladder and tooling; Bad-card HBM signature and the quarantine limit. 4 facts
HalfVulpes on GitHub: Recipes, troubleshooting. Gen3 negative-result controlled study; PCIe Gen3/Gen4 blocked by OTP fuses. 4 facts
ir1ka on GitHub: Unlocks and drivers, community. Upstream ECC branch history: the author’s own measurements; cmpunlocker ECC is three patches. 4 facts
Johnnydaboy on GitHub: Platforms, community. Proxmox dual-CMP in-guest unlock: serialise GSP boots; ECC ‘proven by multiple people’ rests on the reporting patch. 4 facts
JRex286 on GitHub: Unlocks and drivers, card variants, tools. Gen3/Gen4 lock on the 170HX: fuse fingerprint, signed DevInit table; Where GSP decides DRAM ECC: FEATURE_READOUT 0x823814 bit 16. 4 facts
Schaka: Cooling and power, platforms, recipes. Agent serving stack: reasoning_content alias patch is mandatory; PWRBRK# permanent power-brake hazard. 4 facts
vektorprime on GitHub: Community, tools, troubleshooting. Fork ecosystem map: who carries what beyond upstream cmpunlocker; Upstream and fork status as of late October 2026: v0.5 probably final. 4 facts
WU-CVGL on GitHub: Unlocks and drivers, recipes, troubleshooting. Bare-metal static-BAR1 P2P recipes and measured bandwidth; cmpunlocker ECC is three patches. 4 facts
abobasixseven: Platforms, unlocks and drivers. Four PLM registers must be opened, not one; PCIe Gen2 software unlock. 3 facts
asm64-hooligan on GitHub: Cooling and power, community, unlocks and drivers. 170tune HBM overclock: two ceilings; Fork ecosystem map: who carries what beyond upstream cmpunlocker. 3 facts
CurtisAccelerate on GitHub: Recipes. Morrowmake DeepSeek V4.1 Flash recipe v1.0.0 on 4 cards; DeepSeek V4.1 Flash recipe landscape for CMP 170HX and community numbers. 3 facts
aikitoria on GitHub: Recipes, platforms. Bare-metal static-BAR1 P2P recipes and measured bandwidth. 2 facts
buliaoyin on GitHub: Troubleshooting, community. Fork ecosystem map: who carries what beyond upstream cmpunlocker. 2 facts
canoziia on GitHub: Troubleshooting, community. Fork ecosystem map: who carries what beyond upstream cmpunlocker. 2 facts
duggasco on GitHub: Troubleshooting, community. Fork ecosystem map: who carries what beyond upstream cmpunlocker. 2 facts
evanrodgers on GitHub: Tools, card variants. 10 GB card ceilings: 40 GB typical, 48 GB on about 20 %, 128 SM via RIR. 2 facts
Hawk132000: Platforms, unlocks and drivers. bayley fork: static BAR1 P2P on 8 GPUs behind two switches, 56/56 pairs. 2 facts
jeonghyeon00 on GitHub: Community, unlocks and drivers. HBM overclock lineage across cmpunlocker forks. 2 facts
kprinssu on GitHub: Community, unlocks and drivers. Fork ecosystem map: who carries what beyond upstream cmpunlocker. 2 facts
moriyasujapan on GitHub: Recipes, community. Fork ecosystem map: who carries what beyond upstream cmpunlocker. 2 facts
noonghunna on GitHub: Troubleshooting, platforms. Closest analog to in-VM BAR1 P2P. 2 facts
Panchovix on GitHub: Tools. Panchovix VIX VBIOS loader and editor on CMP: in-memory VBIOS at GSP load. 2 facts
sekkit: Recipes, unlocks and drivers. Bare-metal static-BAR1 P2P recipes and measured bandwidth. 2 facts
tehroot on GitHub: Card variants, community. Gen3/Gen4 lock on the 170HX: fuse fingerprint, signed DevInit table. 2 facts
18gogogo on GitHub: Platforms. 1 fact
46cv8 on GitHub: Recipes. 1 fact
aboba: Unlocks and drivers. HBM overclock lineage across cmpunlocker forks. 1 fact
aiultrayang: Unlocks and drivers. VBIOS 92.00.6D cross-flash: what it buys, and why not inside a hypervisor guest. 1 fact
AlexBrandes: Recipes. Landscape of 2-card Qwen3.8-Flash-Next recipes on CMP 170HX: engines. 1 fact
allover326: Recipes. SM80 quant-format compatibility. 1 fact
bendy2: Platforms. PCIe Gen2 software unlock. 1 fact
bluespace3: Recipes. Landscape of 2-card Qwen3.8-Flash-Next recipes on CMP 170HX: engines. 1 fact
bot-lab-21: Recipes. Full GLM-5.3 on 4 CMP + 2 Blackwell. 1 fact
bozocentral: Recipes. Landscape of 2-card Qwen3.8-Flash-Next recipes on CMP 170HX: engines. 1 fact
canada-quant: Recipes. Abliterated GLM-5.3-Flash twin re-quantized to W4A16 for the CMP engine. 1 fact
danvei233 on GitHub: Platforms. 1 fact
dealignai: Recipes. Abliterated GLM-5.3-Flash twin re-quantized to W4A16 for the CMP engine. 1 fact
eejohnso-ops on GitHub: Community. Fork ecosystem map: who carries what beyond upstream cmpunlocker. 1 fact
g4bwy: Recipes. Landscape of 2-card Qwen3.8-Flash-Next recipes on CMP 170HX: engines. 1 fact
halt95: Recipes. Landscape of 2-card Qwen3.8-Flash-Next recipes on CMP 170HX: engines. 1 fact
Hurricane-Computer on GitHub: Platforms. 1 fact
iIIusi0n on GitHub: Recipes. Landscape of 2-card Qwen3.8-Flash-Next recipes on CMP 170HX: engines. 1 fact
incoai: Recipes. Abliterated GLM-5.3-Flash twin re-quantized to W4A16 for the CMP engine. 1 fact
JetSki: Unlocks and drivers. 1 fact
JJ48 on GitHub: Community. Fork ecosystem map: who carries what beyond upstream cmpunlocker. 1 fact
kaka86mm: Recipes. MiMo-V2.6-Flash-RL NVFP4 on 4x 170HX: 322/275/170 tok/s single, 1M context. 1 fact
kha84: Card variants. Single-rank vs dual-rank 8 GB cards. 1 fact
kvk-og: Buying. 1 fact
lazymio: Recipes. Full GLM-5.3 engine blockers on sm_80 and the cheaper alternatives. 1 fact
mannydelrio on GitHub: Recipes. Landscape of 2-card Qwen3.8-Flash-Next recipes on CMP 170HX: engines. 1 fact
mazzanfar on GitHub: Platforms. 1 fact
NA-DEGEN-GIRL on GitHub: Tools. Community tools worth knowing for unlocked cards: testsuite, pstated. 1 fact
nekopep on GitHub: Community. ECC ‘proven by multiple people’ rests on the reporting patch. 1 fact
NeuroSenko on GitHub: Community. Fork ecosystem map: who carries what beyond upstream cmpunlocker. 1 fact
nguyenthimy2022kg-alt: Recipes. Landscape of 2-card Qwen3.8-Flash-Next recipes on CMP 170HX: engines. 1 fact
niconiconi: Community. Card history, press coverage and academic validation. 1 fact
nvlight: Tools. Community tools worth knowing for unlocked cards: testsuite, pstated. 1 fact
phankhiem2329-stack on GitHub: Community. 1 fact
polyuij42: Unlocks and drivers. VBIOS 92.00.6D cross-flash: what it buys, and why not inside a hypervisor guest. 1 fact
pqhaz: Recipes. Abliterating DeepSeek V4.1 Flash on a CMP recipe. 1 fact
promisezackr: Recipes. Pipeline-parallel beats tensor-parallel on the no-NVLink fabric. 1 fact
rovangju on GitHub: Community. Fork ecosystem map: who carries what beyond upstream cmpunlocker. 1 fact
sasha0552 on GitHub: Tools. Community tools worth knowing for unlocked cards: testsuite, pstated. 1 fact
sayhellotojungle on GitHub: Recipes. Bare-metal static-BAR1 P2P recipes and measured bandwidth. 1 fact
snapo: Card variants. Single-rank vs dual-rank 8 GB cards. 1 fact
studebaker8: Platforms. PCIe Gen2 software unlock. 1 fact
superXman1993: Community. 1 fact
tlswotj on GitHub: Troubleshooting. 1 fact
TnzGit on GitHub: Community. 1 fact
vilemooni: Tools. QC acceptance ladder and tooling. 1 fact
wanghar on GitHub: Card variants. 8 GB vs 10 GB: the only RIR difference is slot availability. 1 fact
wgkbfcp on GitHub: Recipes. DeepSeek V4.1 Flash recipe landscape for CMP 170HX and community numbers. 1 fact
XinRanZh on GitHub: Community. 1 fact
y3tiCrypto on GitHub: Community. Fork ecosystem map: who carries what beyond upstream cmpunlocker. 1 fact
zsyonyc on GitHub: Community. ECC ‘proven by multiple people’ rests on the reporting patch. 1 fact

Projects

54 repositories the community builds on, credited to their GitHub owners and to whoever posted them on the Discord.

Core

344303947/CMP-170HX-64GB-Unlocked-VBIOS: A single 1 MB ROM file presented as the first public standalone 64 GB HBM2e VBIOS dump for the CMP 170HX (device 10DE 20C2, subsystem 10DE 1585), by @344303947, posted by @supasnipa360.
ahnguyen17/ga100-fuse-re: Read-only reverse engineering of the GA100 fuse architecture on CMP 170HX from one two-card rig (ASUS Z10PE-D8 WS, C612/Haswell-EP), by @ahnguyen17, posted by @exile_1ck (eXile).
amoghmunikote/170th-Street: GitBook source of ‘170th Street’, the community documentation site for the CMP 170HX written by Amogh Munikote, the author of amoghmunikote/cmpunlocker, by @amoghmunikote, posted by @markshowa.
amoghmunikote/cmpunlocker: Mainline unlock for the NVIDIA CMP 170HX: an installer plus 14 patches against NVIDIA’s open-gpu-kernel-modules that rebuilds nvidia.ko and friends to restore memory geometry, by @amoghmunikote, posted by @68×0 (68x), @_j3ns (Amogh), @arabus_74064 (arabus) and 6 more.
asm64-hooligan/cmpunlocker: The ‘asm64 lineage’ fork of amoghmunikote/cmpunlocker: patched nvidia-open 610.43.03 / 610.43.02 kernel modules that use the SEC2 Booter trick to open PLM gates during GSP bootstrap and restore full, by @asm64-hooligan, posted by @rh3dai (RH-3D).
Consensus-Protocol/cmp170hx: A 55-page, roughly 278,000-word MkDocs reference wiki on the CMP 170HX (GA100): silicon, fuses, VBIOS, the SEC2 Booter ROP unlock, cmpunlocker’s six driver patches and branches, by @Consensus-Protocol, posted by @rrickk420 (Rickcal), @zagitta.
dartraiden/NVIDIA-patcher: Long-running patched-driver project (bilingual EN/RU) that makes NVIDIA’s stock drivers bind mining cards as display/3D-capable GPUs: P106/P104/P102, by @dartraiden, posted by @xeax.
evanrodgers/cmp-170hx-rir-burn-util: Self-described cmp170hx-rir-gen3: an all-in-one operator tool for the 10 GB CMP 170HX (10de:2082) that performs a permanent one-shot RIR fuse burn (clears the DISABLE_SW_OVERRIDE gate) and then, by @evanrodgers, posted by @hezz1891.
justin-kk-coder/CMP-170HX-64GB-Unlocked-VBIOS: A 1 MB VBIOS ROM dump plus a README that presents it as a ’64GB unlocked’ VBIOS for the CMP 170HX (device 10DE 20C2, subsystem 10DE 1585) with nvflash instructions for Windows and Linux, by @justin-kk-coder, posted by @supasnipa360.
lixiangwuxian/170hx-rir-tool: Standalone C tool (ga100_rir.c, about 930 lines, claims no NVIDIA source) that mmaps BAR0 via /sys/bus/pci/devices/<bdf>/resource0 and talks to the GA100 fuse block at 0x820000, by @lixiangwuxian, posted by @btakk_ak47, @dale063629 (大乐), @ghostinthedatacenter_87281 (GhostInTheDatacenter) and 4 more.
lixiangwuxian/170hx_tips: Chinese/English notes on unlocking the 10 GB CMP 170HX variant (PCI 10de:2082; 10de:20c2 is the 8 GB card), by @lixiangwuxian, posted by @catcat0458 (catcat), @lxtend (lxwx).
mintoriakamoto/CMP100-210xFULLTENSORSxGEN3x1: This is cmp100-unlock 1.1.1 (github.com/mintoriakamoto/cmp100-unlock 301-redirects to this repo name): a boot-time, volatile Tensor-core and PCIe Gen3 unlock for the NVIDIA CMP 100-210, by @mintoriakamoto, posted by @jma1753 (JMA1976).
Morrowmake/cmpunlocker: Fork of amoghmunikote/cmpunlocker (tracks upstream master, base 6c442ee) that adds working PCIe peer-to-peer between CMP 170HX cards, by @Morrowmake, posted by @mountains1605 (Mountains), @replacedriver47, @trustinter (Trust Inter).
Morrowmake/vllm-cmp170hx: The CMP 170HX serving engine the community’s vLLM recipes pin: not a classic fork but a `vllm-sm80` plugin (Python package vllm_sm80 with sm_80 kernels, attention backends, by @Morrowmake, posted by @jjb01957 (jjb), @voidgiraffe (Morrowmake).
Panchovix/nvidia-windows-vbios-vix-loader: Loads a modified NVIDIA VBIOS on Windows by storing it in the GPU driver’s registry override key, without ever flashing the card’s SPI chip, then applying it on the next Windows restart, by @Panchovix, posted by @d3m3n7r.
ropuls/d2978531ba3a1a183e70f62f3de2eb84: A GitHub gist (single file cmp-170hx-dual-setup.md, titled “2x CMP 170HX for LLM inference: unlock, 74 SM, cross-root P2P, 300 W VBIOS, undervolt, serving”) by Discord member @ropuls, by @ropuls, posted by @68×0 (68x), @mountains1605 (Mountains), @gnasherrr.
vektorprime/cmpunlocker: Fork of amoghmunikote/cmpunlocker whose master is identical to upstream master as of 2026-10-05 (17535a0; upstream has since only added 615.71.10 to driver/VERSION), by @vektorprime, posted by @rh3dai (RH-3D).

Recipes

alanoo81/deepseek-v41-flash-4x-cmp170hx: A fidelity-first configuration of GNmac-lab/deepseek-v41-flash-1017-cmp170hx (release 1.2.0 at 2ad5741ab1) for DeepSeek-V4.1-Flash-1017 on four CMP 170HX 64 GB cards at TP4 over PCIe Gen2 x16 with, by @alanoo81, posted by @alanoo.
alexvnesta/glm53-flash-cmp170hx-tp: Runs GLM-5.3-Flash (320B total, about 18B active MoE) entirely in HBM on two unlocked CMP 170HX 64 GB cards: a pinned, patched ExLlamaV3 with tensor parallel over NCCL/GPU P2P, by @alexvnesta, posted by @nestaa51.
anoane/09-deepseek-v41-flash-1017-exl3-tp5-1m: Recipe 09 in anoane’s numbered series (01-proxmox-dual-gpu-passthrough is also in links.json): DeepSeek-V4.1-Flash-1017 quantized to EXL3 3.5 bpw by LibertAI, by @anoane, posted by @anon33anon.
arvkmr/vllm-170hx-glm5: Serving recipe and patch set for the full 78-layer GLM-5.3 DSA MoE (GLM-5.3-Int4-Int8Mix-AWQ-g64, not the Flash variant) with the incoai/GLM-5.3-DFlash2 speculative drafter on 8, by @arvkmr, posted by @reala7096 (reala).
clee/qwen3.8-flash-next-3x170hx: A working recipe for serving Intel/Qwen3.8-Flash-Next-W4A16-AutoRound (Qwen4Exp MoE) on three CMP 170HX cards unlocked to 64 GiB, pipeline-parallel 3 with no GPU P2P, by @clee, posted by @c133 (clee), @mountains1605 (Mountains).
CurtisAccelerate/DeepSeekv41-Flash-1017-EXL3-CMP170HX: Weight-selection recipe plus benchmark record for serving DeepSeek V4.1 Flash 1017 on 4x CMP 170HX (64 GB each, sm_80, tensor parallel 4, 140 W, PCIe Gen2 x16, dual-socket Broadwell, no NVLink), by @CurtisAccelerate, posted by @anon33anon, @peter643.
DUFRelic/qwen38-flash-next-recipe: A single-README production recipe for Qwen3.8-Flash-Next on 2 x CMP 170HX (8 GB SKU unlocked to 64 GiB, by @DUFRelic, posted by @dufrelic.
ElectroHobo/mimo-v26-flash-cmp170hx-recipe: Community recipe that serves XiaomiMiMo’s MiMo-V2.6-Flash-RL (309B MoE, 15B active; text, image, by @ElectroHobo, posted by @nathang6866 (ElectroHobo).
ElectroHobo/qwen38-flash-next-3bpw-1x-cmp170hx-recipe: Serves Qwen3.8-Flash-Next (125B MoE) on one CMP 170HX exposing 64 GiB with the full 262,144-token context, by @ElectroHobo, posted by @nathang6866 (ElectroHobo).
ElectroHobo/qwen38-flash-next-w4a16-2x-cmp170hx-recipe: Serves halt95/Qwen3.8-Flash-Next-W4A16-Merlin (INT4 AutoRound experts) on two CMP 170HX cards at 64 GiB each with Morrowmake’s vllm-cmp170hx engine and release image, by @ElectroHobo, posted by @galygious2, @nathang6866 (ElectroHobo).
ElectroHobo/qwen38-flash-next-w4a16-cmp170hx-recipe: This URL is the pre-0.2.0 name of ElectroHobo/qwen38-flash-next-w4a16-2x-cmp170hx-recipe (GitHub answers HTTP 301 to the new name; renamed 2026-10-24), by @ElectroHobo, posted by @nathang6866 (ElectroHobo).
Flun/glm53-flash-cmp170hx-exl3: Full-VRAM serving stack for GLM-5.3-Flash on 2x 64 GB CMP 170HX (SM80, no P2P, host-bounce D2D) built on turboderp’s ExLlamaV3 1.5.4 and the turboderp/GLM-5.3-Flash-exl3 3.05bpw checkpoint, by @Flun, posted by @wingrimez_40467 (Wingrime).
GNmac-lab/deepseek-v41-flash-1017-cmp170hx: Serves DeepSeek-V4.1-Flash-1017 on four CMP 170HX cards (8 GB variant unlocked to 64 GiB, 74 SMs) with vLLM plus a patch stack for sm_80: TP4 by default or PP4 for x1/x4 links, by @GNmac-lab, posted by @greencandle9753 (greencandle).
iIIusi0n/qwen3.8-flash-next-cmp170hx: Single-card vLLM recipe that runs Qwen3.8 Flash Next (AutoRound 3 bpw + MTP checkpoint published by klee100 on Hugging Face) on one 64 GiB SM80 GPU, by @iIIusi0n, posted by @c133 (clee).
mannydelrio/qwen38-flash-next-2x-cmp170hx: Serves Qwen3.8-Flash-Next-Uncensored (512-expert MoE, AWQ W4A16 g32 by leoncca, from orcarouter’s refusal-removed model) on two 8 GB-variant CMP 170HX (PCI ID 10de:20c2) unlocked to 64 GB each, by @mannydelrio, posted by @mrmannyatx (MrManny).
Morrowmake/dsv41-flash-cmp170hx-recipe: Serving recipe that runs DeepSeek-V4.1-Flash-1017 (262,144-token context) on four unlocked CMP 170HX cards behind an OpenAI-compatible API with tool calls, reasoning, by @Morrowmake, posted by @lia_oc (_L_), @mountains1605 (Mountains).
Morrowmake/glm53-flash-cmp170hx-recipe: Serving recipe that runs GLM-5.3-Flash (320B MoE, W4A16, 262,144-token context) on four unlocked CMP 170HX cards (64 GiB each) behind an OpenAI-compatible API with tool calls, reasoning, by @Morrowmake, posted by @jjb01957 (jjb), @mountains1605 (Mountains), @ob88594.
vektorprime/qwen38-flash-next-pp2: Build recipe and vLLM patch set that serves cyankiwi/Qwen3.8-Flash-Next-AWQ-INT4 on 2x CMP 170HX 64G (the unlocked 8 GB card) with –pipeline-parallel-size 2 plus PLE CPU offload, by @vektorprime, posted by @vektor_prime (Ryko).
wgkbfcp/deepseek-v4.1-flash-1017-cmp170hx-recipe: A deployment record (README in Chinese) for DeepSeek-V4.1-Flash-1017 on six CMP 170HX 64G cards at TP6 / PP1 with the DSpark drafter at K=5, by @wgkbfcp, posted by @dale063629 (大乐).
Zanooda/deepseek-v4.1-flash-cmp170hx: Complete serving recipe for DeepSeek-V4.1-Flash-1017 on eight 64 GB CMP 170HX (GA100 sm_80, PCIe Gen2 x4) with vLLM pipeline parallel 8, by @Zanooda, posted by @gnasherrr.

Tools

cachenetics/170tune: Userspace clock, undervolt, HBM memory-clock (NDIV), DRAM-timing and refresh tuning harness for the CMP 170HX, applied live through BAR0 register writes with no driver rebuild or VBIOS flash, by @cachenetics, posted by @drew_2 (drew-3), @rh3dai (RH-3D).
cachenetics/gds-nvme-patch: Kernel NVMe-driver patch plus installer that restores NVIDIA GPUDirect Storage (cuFile / nvidia-fs direct NVMe-to-VRAM DMA) on Linux kernels 6.18 and newer, by @cachenetics, posted by @mountains1605 (Mountains).
dev-boz/cmp170hx-testsuite: Bash plus CUDA acceptance test suite for unlocked CMP 170HX cards (64 GB on the 8 GB board 10de:20c2, 40 GB on the 10 GB board 10de:2082), written to catch HBM faults that pass light tests, by @dev-boz, posted by @boz._ (B-st).
DingoOz/llm-visuals: A Rust/ratatui live terminal dashboard for a local inference server: it auto-detects llama-server, ollama, vLLM, SGLang, Strata, Maya, NInfer and TensorFold, reads their counters, by @DingoOz, posted by @trustinter (Trust Inter).
HalfVulpes/Deepinfer-SM80: Ports the ideas of DeepSeek’s DeepInfer (Blackwell tcgen05/TMEM/CuteDSL kernels) to Ampere SM80/SM86 as two vLLM patches plus 30 drop-in Python/Triton/TileLang modules, by @HalfVulpes, posted by @reala7096 (reala), @segmond.
local-inference-lab/llm-inference-bench: General LLM serving benchmark: llm_decode_bench.py (Rich TUI) measures prefill and sustained/burst decode throughput over a concurrency x context matrix against vLLM, SGLang, by @local-inference-lab, posted by @mountains1605 (Mountains), @nestaa51.
matstatman/gpcprobe: Read-only out-of-tree Linux kernel module (gpcprobe.ko) plus two Python frontends for GA100 cards: it ioremaps the first 16 MiB of each NVIDIA GPU’s BAR0 while the proprietary driver stays bound, by @matstatman, posted by @matstatman.
matstatman/hbmmon: nvtop-style terminal monitor for the CMP 170HX (GA100) that derives the HBM topology from live hardware rather than tables, by @matstatman, posted by @matstatman.
ReinForce-II/mmapeak: CUDA micro-benchmark that measures peak dense tensor-core MMA throughput per data type (INT4, INT8, FP16, BF16, TF32, by @ReinForce-II, posted by @jimtjames (gym).
sasha0552/nvidia-pstated: A daemon (sasha0552, author of other p-state/vLLM tooling popular in mining-to-LLM circles) that automatically manages NVIDIA GPU performance states to cut idle power draw and can drive external fans, by @sasha0552, posted by @d3m3n7r.
vektorprime/170tune: Fork of cachenetics/170tune (author studebaker8, copyright Cachenetics; also mirrored at studebaker8/170tune), the userspace tuning and qualification harness for the CMP 170HX, by @vektorprime, posted by @rh3dai (RH-3D).
zzc0721/torch-performance-test-data: PyTorch square-GEMM throughput benchmark (test.py, benchmark version 2.0) with a community leaderboard (database.md plus perf.svcfusion.com) for FP32, TF32, FP16, BF16 and experimental FP8, by @zzc0721, posted by @jimtjames (gym).

Platforms

adrienbrault/cmp-170hx-duct: Parametric, 3D-printable ducting that cools two passively cooled CMP 170HX cards with a single 120 mm Arctic P12 Pro PST CO fan, by @adrienbrault, posted by @adrienbrault.
anoane/01-proxmox-dual-gpu-passthrough: Host-side Proxmox VE 9.2 recipe that passes an RTX PRO 6000 Blackwell Workstation (96 GB, sm_120, 10de:2bb1) and an 8 GB CMP 170HX unlocked to 64 GB (10de:20c2, by @anoane, posted by @logy0 (Logy).
dkpoulsen/170hx-fanduct: Parametric OpenSCAD fan duct (one file, 3x170hx_duct.scad) that lofts a rectangular 120 x 240 mm fan opening into N stacked GPU inlet connectors for CMP 170HX cards in multi-GPU compute or mining, by @dkpoulsen, posted by @c133 (clee).
NA-DEGEN-GIRL/wrx90-bios-mod: An experimental, AI-assisted BIOS modification that adds the PCIE X8X8 Mode bifurcation option (and its early PCIe lane-configuration logic) to the ASUS Pro WS WRX90E-SAGE SE on BIOS 1317, by @NA-DEGEN-GIRL, posted by @stcord (ST).
PixelML/club-170hx: Club CMP 170HX is PixelML’s community knowledge base for the card: hardware and buying notes, a Proxmox installation baseline (VFIO guest and later bare-metal with a GPU LXC), by @PixelML, posted by @c133 (clee), @djpandaman (mrpandagod).