Known bad combinations
From CMP Wiki
More actions
Combinations that failed for at least one person, with what happened and what worked instead. A single report does not prove a part is bad for everyone; it tells you what to try first when you hit the same thing. Each row links to the page with the full story. If something here works for you, say so on the talk page with your setup.
Drivers and unlock builds
| Combination | What happens | Use instead | Status | Source |
|---|---|---|---|---|
| nvidia-open 615.78.08 + cmpunlocker | The unlock fails: PCIe 2.0 training fails at boot and the cards are missing from nvidia-smi; the v0.5 ECC install errors out; the Booter image changed in this release. Upstream lists the version as supported. | 615.71.09 (610.57.04 as fallback) | Proven | Source: @drew_2, 2026-10-10 Source: @zagitta, 2026-10-17 Source: @catchyuser, 2026-10-20 Source: @.henrykrinkle, 2026-10-20 |
| cmpunlocker v0.2 | Hands out a protected 141 MB region at the top of memory; long prompts then crash the second card with Xid 31 | v0.5 or newer | Reported | Source: repo github.com/mannydelrio/qwen38-flash-next-2x-cmp170hx by mannydelrio on GitHub, posted by @mrmannyatx, 2026-10-06 |
| cmpunlocker v0.4 to v0.5 upgrade (one host) | Qwen 3.8 27B decode fell from 60-80 tok/s to 35 tok/s; ECC suspected, unresolved | Not known | Reported | Source: @mowgli8628, 2026-10-18 |
| Morrowmake cmpunlocker fork + ECC | The fork carries P2P but removes the ECC patches | Upstream for ECC; merge P2P by hand if you need both | Reported | Source: repo github.com/Morrowmake/cmpunlocker by Morrowmake on GitHub, posted by @trustinter, 2026-10-07 |
| 170tune + asm64-hooligan cmpunlocker | Preflight fails: PLL PLM 0xBADF2010; that unlock opens only the broadcast FBPA PLL window, not the unicast window 170tune uses
|
vektorprime fbpa-pll-170tune + 170tune fbpa-first-active, or bake NDIV at build time
|
Reported | Source: repo github.com/asm64-hooligan/cmpunlocker by asm64-hooligan on GitHub, posted by @rh3dai, 2026-10-04 Source: @rh3dai, 2026-10-04 |
170tune + a driver built with --mclk-ndiv
|
170tune refuses to tune | A stock-memory-clock build | Reported | Source: repo github.com/vektorprime/170tune by vektorprime on GitHub, posted by @rh3dai, 2026-10-04 |
| cachenetics/170tune latest + amoghmunikote master, and the vektorprime pair (one owner's cards) | Preflight fails with "FBPA privilege masks are closed" on every card | Unresolved; one other owner runs the odd card on plain v0.5 without 170tune | Reported | Source: @rh3dai, 2026-10-11 Source: @drew_2, 2026-10-12 |
P2P registry string in modprobe.d on a host with other NVIDIA GPUs
|
The options are global: they also force the other cards (an RTX 4080 Super in the report) to Gen2 | Only on CMP-only hosts | Reported | Source: @ybryany, 2026-10-14 Source: @68x0, 2026-10-14 |
| gpu_unlock (consumer tensor unlock) + cmpunlocker on kernel 7.2.11, with other NVIDIA cards in the same box | cmpunlocker link retraining broke until the kernel and driver were rolled back; the owner suspects gpu_unlock was "possibly not aiming just for 3090 cards". Another member says gpu_unlock runs before the NVIDIA driver loads, is gated by PCI device ID and should just work beside cmpunlocker. | Roll back the kernel and driver | Contested | Source: @lycorascence, 2026-10-23 Source: @68x0, 2026-10-26 |
| Panchovix nvidia-linux-vbios-vix-loader + CMP 170HX | The loader refuses unknown PCI IDs; 0x20C2 and 0x2082 are not in its list and GA100 is excluded from its relocation flows | None for the 170HX as shipped | Reported | Source: repo github.com/Panchovix/nvidia-linux-vbios-vix-loader by Panchovix on GitHub, posted by @68x0, 2026-10-17 |
| NVIDIA-patcher "NVENC" flavour (Windows) + CMP 170HX | Crashes apps on cards without hardware NVENC; NVENC/NVDEC is disabled on the 170HX | The regular flavour | Reported | Source: repo github.com/dartraiden/NVIDIA-patcher by dartraiden on GitHub, posted by @xeax, 2026-10-27 |
Driver installed with --no-install-libglvnd + memtest_vulkan
|
Vulkan loader ICD error, memtest_vulkan will not start | Reinstall with libglvnd | Reported | Source: @alanoo, 2026-10-05 |
| "CMP-170HX-64GB-Unlocked-VBIOS" GitHub repos | The VBIOS file did not differ from stock when compared; the unlock makes no VBIOS change. A bad flash can leave a card unbootable. | Do not flash anything from them | Warning | Source: @xanthefin, 2026-10-18 Source: @c133, 2026-10-18 |
| nvidia-open 615.71.09 + cmpunlocker in a VM that also holds RTX PRO 6000 (GB202) cards | The unlock works, but any CUDA context wedges UVM (krcWatchdog: GPU is probably locked, processes stuck in uvm_va_space_destroy); a known NVIDIA 615.71.09 issue
|
The unlock ported to 580.178.04, or keep those cards off the shared 615 driver | Warning | Source: @darkmaniac7, 2026-10-21 (lab notes, darkmaniac7/cmp170hx-esxi) Source: @darkmaniac7, 2026-10-25 (lab notes, darkmaniac7/cmp170hx-esxi) |
vektorprime fbpa-pll-170tune, admunch888 --mclk-ndiv, or the "NDIV 45 + refresh 24" tune, on a card with FBPA0 fused off
|
The FBPA0 PLL writes stop the card from booting (GSP booter 0xffff); two of eight cards in one box
|
Read the fuse map first; the Morrowmake lineage (no PLL writes) runs those cards | Warning | Source: @darkmaniac7, 2026-10-27 (lab notes) |
RMForceStaticBar1 / RMPcieP2PType in NVreg_RegistryDwords, GSP 610.43.02
|
NV_ERR_INVALID_REGISTRY_KEY; the protected region stays up until a cold power-off
|
Pass those keys the way your fork documents | Reported | Source: @darkmaniac7, 2026-10-19 (lab notes) Source: repo github.com/LingzheZhao/cmpunlocker by LingzheZhao on GitHub |
Details: Driver and unlock failures.
Kernels, distributions and desktops
| Combination | What happens | Use instead | Status | Source |
|---|---|---|---|---|
| Secure Boot enabled | The installer refuses; the patched modules are unsigned | Secure Boot off | Reported | Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub, posted by @rh3dai, 2026-10-08 |
| Kernel upgrade after installing the unlock | The installer removes DKMS nvidia modules, so the patched modules are gone for the new kernel; the card can silently come up at stock size | Re-run the installer after every kernel upgrade, or hold the kernel | Reported | Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub, posted by @rh3dai, 2026-10-08 Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 |
| Mainline kernel swapped in on Ubuntu | The open driver build breaks | The distribution kernel | Reported | Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @rrickk420, 2026-10-09 |
| Kernel 7.2.11 + cmpunlocker link retraining (10 GB card after RIR burn) | Link retraining broke | Rolled back kernel and driver | Reported | Source: @lycorascence, 2026-10-23 |
Newer kernel with CONFIG_IO_STRICT_DEVMEM + 170hx-rir-tool
|
The tool cannot reach the registers it needs | LTS kernel with the oldest driver cmpunlocker supports, plus iomem=relaxed
|
Reported | Source: @lycorascence, 2026-10-23 |
| LTS kernel + unbinding the card under KDE/sddm | Unbind and rebind hard-crashed sddm (it worked on the newer kernel) | Disable the desktop and work over SSH | Reported | Source: @lycorascence, 2026-10-23 Source: @intervitens, 2026-10-23 |
| Wayland or Xorg desktop + a headless CMP you need to unbind | The desktop grabs the card and blocks the unbind; excluding it broke the KDE login | A headless host (disable the GUI, SSH in from another machine). A live disk alone did not get the RIR tool past the DEBUGCTRL latch for the same owner. | Reported | Source: @lycorascence, 2026-10-17 Source: @lycorascence, 2026-10-20 Source: @intervitens, 2026-10-23 |
| Gaming on Linux vs Windows | About 20 fps on Linux against 60 fps on Windows on the same card; something in the open driver is suspected of throttling | Not known | Reported | Source: @bigptoughneigh, 2026-10-17 |
Engines and recipes
| Combination | What happens | Use instead | Status | Source |
|---|---|---|---|---|
| Morrowmake GLM-5.3-Flash recipe 1.7.0 / 1.7.1 | Endless repetition, nonsense or Chinese text mid-reasoning | 1.7.2 or later (1.6.0 was unaffected) | Proven | Source: @cmdrwaka324, 2026-10-08 |
| Morrowmake GLM-5.3-Flash runtime before the 2026-10-06 release | Memory leak in a prefill step | The 2026-10-06 release or later | Reported | Source: @spectakkular, 2026-10-06 Source: @voidgiraffe, 2026-10-06 |
| tokenspeed engine + CMP 170HX | Refuses to start: needs compute capability 8.9 or newer; a PR adding SM80 support was closed as won't-fix; no pipeline parallelism | vLLM or llama.cpp | Reported | Source: @mountains1605, 2026-10-05 Source: @c133, 2026-10-05 |
vLLM at gpu-memory-utilization 0.95
|
Card crash; only 64733 of 65052 MB is really available | 0.90 | Reported | Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @rrickk420, 2026-10-09 |
| vLLM PP + MTP + prefix caching + mamba alignment | NaN logits and "duct duct duct" output after a few dozen parallel requests | The prefix-caching patch in the recipe repo | Reported | Source: repo github.com/vektorprime/qwen38-flash-next-pp2 by vektorprime on GitHub, posted by @vektor_prime, 2026-10-07 |
| GDS loaders (for example fastsafetensors) + low PCIe width | Null-dereference crash in nvidia-fs or a hang at 0% | Bounded staging (gds_bounded)
|
Reported | Source: repo github.com/cachenetics/gds-nvme-patch by cachenetics on GitHub, posted by @mountains1605, 2026-10-20 |
Docker --gpus all after a crash
|
Devices go missing | Bind an explicit device list (--gpus '"device=0,1,2,3"')
|
Reported | Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 |
| Morrowmake PCIe P2P custom all-reduce with a stale compile lock | All workers hang at init; on Gen2 TP4 it is also slower once it runs (about 40% less decode) | Morrowmake vllm-cmp170hx commit f5ba3755, own compile cache per instance; NCCL only at TP4 | Proven | Source: repo github.com/Morrowmake/vllm-cmp170hx by Morrowmake on GitHub, posted by @voidgiraffe, 2026-10-15 Source: @darkmaniac7, 2026-10-28 (lab notes) |
| vllm-backport engine, DeepSeek-V3.2-style model code, TP2 x PP4 | Pipeline hand-off sends an un-reduced partial sum: degenerate output, lower recall (vLLM PR #60477 is the same bug) | The one-line worker fix, or PP8 / TP8 | Proven | Source: @darkmaniac7, 2026-10-28 (lab notes) |
| vllm-backport 753B GLM-5.3 build + MTP + schema-constrained output | HTTP 500 per request (grammar rejects the draft tokens) | No speculation for schema traffic | Proven | Source: @darkmaniac7, 2026-10-28 (lab notes) |
| GNmac DeepSeek V4.1 Flash TP4 port at 75-92k tokens | Long-context verbatim copying corrupted (85% of paths wrong) | The Morrowmake DeepSeek recipe (0% wrong with the same expert files) | Proven | Source: repo github.com/GNmac-lab/deepseek-v41-flash-1017-cmp170hx by GNmac-lab on GitHub, posted by @greencandle9753, 2026-10-21 Source: @darkmaniac7, 2026-10-29 (lab notes) |
| MTP on PP4 in the Morrowmake GLM fork | Draft acceptance 0.27 / 0.11 / 0.08, slower than no drafter | DFlash2 | Proven | Source: @darkmaniac7, 2026-10-19 (lab notes) |
| OneZeroMiner core offset on cards with the stock 250 W VBIOS | "Failed to apply Core offset: an internal driver error"; that VBIOS exposes no V/F offset range | Power limit only | Reported | Source: @darkmaniac7, 2026-10-28 (lab notes) |
Details: Instability under load.
Motherboards and BIOS
| Board | What happens | Workaround | Status | Source |
|---|---|---|---|---|
| ASUS WRX80E | BAR1 stuck at 64 MB | None posted; mailbox P2P at 64 MB | Reported | Source: @trustinter, 2026-10-10 |
| ASUS WRX90 | BAR1 cannot be assigned | None found | Reported | Source: @ghostinthedatacenter_87281, 2026-10-07 |
| ASUS Pro WS WRX80E-SAGE | Asserts PWRBRK# on edge pin B30 and pins the card at about 88 W and 1140 MHz | Kapton tape over pin B30, or a riser that does not route B30 | Reported | Source: repo github.com/Zanooda/deepseek-v4.1-flash-cmp170hx by Zanooda on GitHub, posted by @gnasherrr, 2026-10-17 |
| Minisforum MS-01 + PLX 88096, 4 cards | MMIO is 256 GB, too small for four 64 GB BAR1s | Mailbox P2P (another owner runs 8 cards on MS-01 hosts that way) | Reported | Source: @con3po7016, 2026-10-22 Source: @replacedriver47, 2026-10-22 |
| ASUS P9X79 WS | 64-bit MMIO window too small for a 64 GB BAR | MMIOH size raised from the OS side (parameters not posted) | Reported | Source: @fdelfini, 2026-10-19 |
| ASRock X370-I | Would not POST with a CMP as the only card (no video output) | Add a display GPU or use an iGPU | Reported | Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @rrickk420, 2026-10-09 |
| ASUS X99-A | Gen2 x1 in one slot only, Gen1 x4 in the others (July 2026, with an old retrain tool that hard-coded the PCI address) | Likely fixed by the later tool; not confirmed | Reported | Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @rrickk420, 2026-10-09 |
| Threadripper board, cards on different PCIe root complexes | P2P set up but no traffic arrived; on another 3-card Threadripper host, DMA between roots hung | NCCL_P2P_DISABLE=1, or a local P2P patch for the topology
|
Reported | Source: repo github.com/clee/qwen3.8-flash-next-3x170hx by clee on GitHub, posted by @c133, 2026-10-04 Source: @c133, 2026-10-22 |
| Dual-socket hosts, cards on different sockets | P2P across sockets does not work, according to one member; another says it can be mitigated by attaching the PLX switch to both CPUs. Measured against that: an 8-card dual-EPYC Supermicro H12DSG-O-CPU passed 504 of 504 P2P content checks, cross-socket pairs included, with the Morrowmake P2P build on 615.71.09 (Supermicro H12DSG-O-CPU) | Keep each tensor-parallel group on one socket unless a content check passes across sockets | Contested | Source: @wakeupmbrc1, 2026-10-18 Source: @endsjeremy, 2026-10-18 Source: @darkmaniac7, 2026-10-27 (lab notes) |
| Supermicro H12D-8D + DDR4-2133 4DRx4 LRDIMMs | About 120 corrected host ECC errors per minute | DDR4-3200 2Rx4 RDIMMs | Reported | Source: @fafinet, 2026-10-25 |
| Some Intel hosts | Gen2 never trained across branches and slots, while an AMD host did with no kernel options | An AMD host, or a retrain that fires inside the card's window | Reported | Source: @darkmaniac7, 2026-09-24 (lab notes) Source: repo github.com/rjohny55/cmp170hx-unlocker by rjohny55 on GitHub |
Cards whose PCB revision locks the Gen2 path (FAILED to set OPT_GEN23 in dmesg)
|
64 GB and the SM unlock work, Gen2 does not | None known | Reported | Source: @darkmaniac7, 2026-09-24 (lab notes) Source: repo github.com/rjohny55/cmp170hx-unlocker by rjohny55 on GitHub |
Every BIOS on this list: see also Card not detected or BAR allocation fails and Platforms.
Switches, risers and enclosures
| Combination | What happens | Workaround | Status | Source |
|---|---|---|---|---|
| PLX switch + mailbox P2P + tensor parallel | Xid 13 crashes the inference program | PP, P2P off, or no PLX for TP | Reported | Source: @ob88594, 2026-10-07 |
| PLX 88096 switch firmware without hotplug | BAR addressing cannot be remapped | Switch firmware with hotplug enabled | Reported | Source: @replacedriver47, 2026-10-24 |
| PLX 88096, four cards, two of them at 64 GB BAR1 after MMIOH and BAR tuning, Morrowmake cmpunlocker | P2P transfers never complete, even between the two cards with the 64 GB BAR1; root cause unknown | None confirmed. PixelML's 4-card BAR1 P2P over PLX notebook is the reference people point to; ElectroHobo thinks Morrowmake's vLLM fixes it on supported hardware. See Risers and bifurcation | Reported | Source: @replacedriver47, 2026-10-24 Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @djpandaman, 2026-10-24 Source: @nathang6866, 2026-10-25 |
| Two cards behind one PLX switch, BAR1 resized per card without locking | Kernel resource tree corrupted on every boot | rebar-serialize.patch (same fix as upstream PR #59)
|
Reported | Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 |
| PEX88096 board with SAS cables in the wrong ports | Host link only x8 | Re-cable in order | Reported | Source: @domrockt2, 2026-10-15 |
| PLX board + two independent 750 W PSUs | All four cards and the board lose power under big prefill | Unresolved (see Instability under load) | Reported | Source: @replacedriver47, 2026-10-05 Source: @reala7096, 2026-10-05 |
| Generic passive PCIe riser | The BMC sees an add-in card, not a GPU, and cannot read its temperature | A reputable riser brand, or fan control from the OS | Reported | Source: @fafinet, 2026-10-21 Source: @endsjeremy, 2026-10-21 |
| Razer TB3 enclosure on X299 | Did not work | Native slot or OCuLink | Reported | Source: @_orrorin, 2026-10-18 |
| Dell OptiPlex 7040m (16 GB RAM) + OCuLink eGPU | Host swapped and restarted under LLM load; OCuLink a bit flaky | A host with 32 GB or more | Reported | Source: @globalapathy, 2026-10-18 |
| eGPU dock over USB4 | In use when a card started failing GSP init on every boot; a native slot test was pending | Test in a native PCIe slot | Reported | Source: @mistyashhole, 2026-10-10 |
| BAR1 P2P across root ports with ACS redirect enabled | 100% of a 256 MiB device-to-device copy lost, no error | pci=disable_acs_redir=... for the switch ports
|
Warning | Source: @darkmaniac7, 2026-10-19 (lab notes) Source: repo github.com/lesj0610/cmpunlocker by lesj0610 on GitHub |
| PLX PEX8747/8749 hosts on a stock kernel, several cards with 64 GB BAR1 | Some cards silently get no BAR1 (bridge-window sizing bug) | The two host-kernel patches, pci=hpmmioprefsize=2T, a large MMIO-high window
|
Reported | Source: @darkmaniac7, 2026-09-12 (lab notes) Source: repo github.com/bayley/cmpunlocker by bayley on GitHub |
| Crucial T700 (PCIe 5 NVMe) boot drive on a PCIe-to-M.2 sled with no heatsink, under DeepSeek Engram disk reads | The drive overheated and dropped off the bus; the host would not boot and the journal was lost | Heatsink and fan, a read-rate cap, models on a separate drive | Warning | Source: @darkmaniac7, 2026-10-28 (lab notes) |
Virtual machines
| Combination | What happens | Use instead | Status | Source |
|---|---|---|---|---|
| Proxmox VM with OVMF (UEFI) | RmInitAdapter and SEC2 failures that look like a failed unlock
|
SeaBIOS, q35 | Reported | Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @rrickk420, 2026-10-09 Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 |
| Running the unlock inside the guest | Unreliable with two cards | Unlock on the host, then pass through | Reported | Source: @bismuth67, 2026-10-08 Source: @spiralglide, 2026-10-08 |
| VFIO passthrough on a CachyOS host | Did not work | Not known | Reported | Source: @ybryany, 2026-10-14 |
cmpunlocker-passthrough.service on a host that does not pass cards to VMs
|
Blocks GPU resets | Disable the service | Reported | Source: repo github.com/mannydelrio/qwen38-flash-next-2x-cmp170hx by mannydelrio on GitHub, posted by @mrmannyatx, 2026-10-06 |
| Any VFIO passthrough | No P2P between cards | Bare metal or containers for multi-card TP | Reported | Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 |
| Static BAR1 P2P between CMP cards in an ESXi guest | Init fails, or the first CUDA context wedges UVM (hard VM reset); no CMP P2P works in any hypervisor so far | NCCL_P2P_DISABLE=1 in VMs; P2P on bare metal
|
Warning | Source: @darkmaniac7, 2026-10-19 (lab notes, darkmaniac7/cmp170hx-esxi) Source: @darkmaniac7, 2026-10-21 (lab notes, darkmaniac7/cmp170hx-esxi) |
| Upstream in-guest unlock under QEMU/VFIO | Booter returns 0x15 instead of 0x31; unlock fails
|
Unlock on the host; or a fork that unlocks in the guest | Reported | Source: @darkmaniac7, 2026-10-19 (lab notes) Source: repo github.com/amoghmunikote/cmpunlocker by amoghmunikote on GitHub |
| The FEAT_OVR_ECC_PLM write in a dual-CMP Proxmox guest | Fatal Booter 0x15 on the second card
|
Drop that write; serialise GSP boots | Reported | Source: @darkmaniac7, 2026-09-04 (lab notes) Source: repo github.com/Johnnydaboy/cmpunlocker by Johnnydaboy on GitHub |
bar1-resize-unlock.patch behind VFIO, or on one 8-card switch server
|
BAR2/MMU init failures behind VFIO; the server hung | Build without it (mailbox P2P runs at 64 MB BAR1) | Reported | Source: @darkmaniac7, 2026-10-19 (lab notes) Source: repo github.com/Johnnydaboy/cmpunlocker by Johnnydaboy on GitHub Source: repo github.com/cebeuq/glm53-cmp170hx-recipe by cebeuq on GitHub |
| ESXi passthrough and Resizable BAR | ESXi ignores ReBAR for passthrough (64 MiB BAR1); every VM power-on resets the card's override to 256 MiB; a normal host reboot loses a 64 GiB allocation | A pre-boot UEFI writer plus a guest-side repair hook; see BAR1 64 GB on ESXi | Reported | Source: @darkmaniac7, 2026-10-19 (lab notes, darkmaniac7/cmp170hx-esxi) Source: @darkmaniac7, 2026-10-21 (lab notes, darkmaniac7/cmp170hx-esxi) |
hypervisor.cpuid.v0 = FALSE in a VM with CMP cards
|
Hides the CPUID leaves the driver uses to detect VMware; changes driver behaviour for every GPU in the VM | Leave it at the default | Reported | Source: @darkmaniac7, 2026-10-21 (lab notes, darkmaniac7/cmp170hx-esxi) |
Power parts
| Part or setup | What happened | Use instead | Status | Source |
|---|---|---|---|---|
| One PCIe cable (plus its pigtail) feeding a card's 2x PCIe-to-EPS adapter | Burned a DeepCool PX1000P after about 48 h at 220 W per card; "magic smoke" | Two separate PCIe cables per card, or a direct EPS cable | Warning | Source: @con3po7016, 2026-10-11 |
| 6-pin to 8-pin adapters | One of six melted at 170 W within weeks | Two full 8-pin leads per card, or proper EPS cables | Warning | Source: @neapopa, 2026-10-17 |
| Lian Li 1200G at about 600 W | Thermal shutdown after hours; very slow fan curve | Extra airflow on the PSU, or another unit | Warning | Source: @phoenixone7, 2026-10-05 |
| Refurbished HP 1200 W server PSUs (eBay), 10 GPUs at 160 W | One blew up after about two hours at about 55% load | Test used server PSUs with a CRPS tester first | Warning | Source: @reala7096, 2026-10-20 Source: @endsjeremy, 2026-10-20 |
| ZSX-AMPS server-PSU breakout boards, 3-4 years old | Constant reboots and PSU shutdowns under load | Replaced with X11-AMPS boards | Warning | Source: @finishtilt14, 2026-10-15 |
| 4+4 EPS (CPU) cables | The double-width latch does not close on the card | Trim the latch | Proven | Source: @civiloid, 2026-10-09 Source: @c133, 2026-10-19 |
| 300 W VBIOS | A burned 170HX was reported | Stay on the stock power limit | Warning | Source: @xanthefin, 2026-10-18 |
Details: Power and cabling warnings.
Cards that are a bad buy
| Card | What it is | How to spot it | Status | Source |
|---|---|---|---|---|
| "Brand new, never used" 170HX sold in China | PG199-origin boards repurposed by NVIDIA as 170HX; they unlock to 32 GB only | PCI subsystem 10de:1586, GPU part 20C2-105-A1, board 699-11001-0109-RC1, engineering-sample VBIOS 92.00.6D.00.01
|
Warning | Source: @_arcanum_, 2026-10-13 Source: @boz._., 2026-10-13 Source: @douya4305, 2026-10-13 Source: @wakeupmbrc1, 2026-10-13 |
| A "170HX 8GB" listing at 9999 RMB from Taobao store shop35050154 (a large, well-rated retailer with many cards) | Another member identified the cards as PG199HX-origin (Tesla-derived), not 170HX: "that was the catch" | Read the listing details; check board part and VBIOS before paying 170HX prices | Warning | Source: @spectakkular, 2026-10-11 Source: @dalzyyy, 2026-10-11 |
| Single-rank 8 GB cards | Unlock to 32 GB, not 64 GB | A CUDA memory test past 32 GB | Reported | Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 |
| A "170HX" with a transplanted die | One card turned out to carry a die from the 32 GB A100 variant (recalled from an earlier report) | SM count and memory size after unlock | Reported | Source: @civiloid, 2026-10-13 |
| Cheap "8 GB" cards from China | Often the cards that failed the unlock: HBM errors, GSP will not initialize, engineering samples | Insist on an unlocked-and-tested card | Warning | Source: @bayley_w, 2026-10-13 Source: @ob88594, 2026-10-13 Source: @replacedriver47, 2026-10-13 |
| CMP cards bought through the Superbuy purchasing agent (also via Goofish) | Superbuy has no GPU test bench, and cards bought that way arrived dead | Buy where the card can be tested or returned | Warning | Source: @kellertreppe123, 2026-10-25 |
| Used 10 GB cards | Expected to show up with RIR slots already spent and nothing unlocked | Dump the fuses (gpcprobe or the RIR tool, read only) before you pay | Theory | Source: @dottoreargus, 2026-10-20 |
Failure rate of bought cards Contested
- One of two 8 GB cards was bad and went back.
Source: @ghostinthedatacenter_87281, 2026-10-07 - The cards have a 50% failure rate.
Source: @_____josh_____, 2026-10-08 - "Nowhere near that high."
Source: @lia_oc, 2026-10-08 - 5 of 5 tested cards worked.
Source: @c133, 2026-10-08 - 12 of 12 cards bought early all unlocked.
Source: @bigfoot1144, 2026-10-08 - About 20% of bought cards fail, and fake cards have been reported.
Source: @civiloid, 2026-10-13
Bulk resellers of mining pulls usually do not test. See Buying and Which card do I have.
Source: @spectakkular, 2026-10-08