cmp170hx-testsuite
More actions
Reported 8 GB and 10 GB
cmp170hx-testsuite is an acceptance test suite for unlocked CMP 170HX cards (64 GB on the 8 GB card 10de:20c2, 40 GB on the 10 GB card 10de:2082). It was written to catch HBM faults that pass light tests, most memtest_vulkan runs and every compute test. Each command says what it does, shows the exact command and the pass rule, shows live output, ends with PASS or FAIL, and saves raw logs per card serial and date, so a session can be filmed or shared with a seller. Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09.
Author: @boz._. (GitHub dev-boz), MIT. "i got claude to build a test suite for new cards"; he invites people to add to it or fork it. Source: @boz._., 2026-10-09 Repository: https://github.com/dev-boz/cmp170hx-testsuite
Commands
| Command | Time | Checks | Pass rule |
|---|---|---|---|
cmp-identity |
instant | serial, UUID, VBIOS, memory size, PCIe link, kernel GPU errors | CMP 170HX with full unlocked memory (65536 or 40960 MiB), 0 Xid |
cmp-hbm-test [min] |
10 min (5 is enough) | all free VRAM, walking-bit pattern alternating with full-bandwidth random data, every word verified | 0 errors |
cmp-tensor-test [s] |
5 min | INT8 and FP16 tensor-core matmuls on every SM against an exact reference | 0 errors, every SM used |
cmp-sm-test |
under 1 min | fixed workload on every SM, bit-compared with a known-good baseline | bit-identical |
cmp-ai-test |
2 to 4 min | real llama.cpp inference plus every weight read back from VRAM | expected answer, no NaN, 0 bytes differ |
cmp-kernel-check |
instant | NVIDIA Xid reports since the first test today | 0 Xid |
cmp-summary |
instant | table of today's results for this card | |
cmp-hbm-map [min] |
5 min | diagnostic: maps HBM errors to fixed 2 MiB physical blocks | |
cmp-telemetry |
live | temperatures, power, load, VRAM, clocks, blower rpm, Xid count |
Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09.
Install
Needs Linux with a driver that runs the card unlocked, the CUDA toolkit (nvcc, sm_80), and read access to the kernel log. cmp-ai-test also needs llama.cpp built with CUDA and a GGUF that fits on one card. Developed on driver 610.57.04 (open modules with an unlock patch set), CUDA 12.8, Linux 7.0, an X399 Threadripper board. Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09.
git clone https://github.com/dev-boz/cmp170hx-testsuite && cd cmp170hx-testsuite
LLAMA_DIR=~/llama.cpp ./install.sh # builds bin/, links cmp-* into ~/.local/bin; omit LLAMA_DIR to skip the AI test
cp config.example.sh config.sh # optional: model path, blower tach mapping, output folders
./install.sh remove # removes the links
Usage
One card at a time; cap the power (sudo nvidia-smi -i <index> -pl 180, resets every boot) and run the blower at full speed. GPU tests run under scripts/guard.sh, which stops them at 72 C core or 85 C HBM (GPU_LIMIT, MEM_LIMIT). Suggested order, about 40 minutes: cmp-identity; memtest_vulkan for 5 minutes right after a cold boot; gpu_burn -tc -m 95% 300; cmp-hbm-test 5, cmp-tensor-test, cmp-sm-test, cmp-ai-test; cmp-kernel-check, cmp-summary. docs/filming.md has a run sheet for filming it in one take. Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09. The full sequence with the other tools is on Testing and burn-in.
The bad card it was built on
Warning Unlocked to 64 GB, 180 W, normal temperatures, faulty in every slot and on every cable: whole 128-byte blocks read back as zeros or unrelated data from the first second of each run; gpu-burn reported 0.6 to 1.9 billion errors; memtest_vulkan failed only right after a cold boot; compute was clean. Under load its firmware crashed (Xid 119, Xid 1, Xid 154 or Xid 62) and only a 30-second power-off recovered it. cmp-hbm-map showed 24 percent of VRAM failing within 58 seconds, with errors following launch timing rather than location: a load-transient fault that block quarantine cannot fix. Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09.
Caveats
- One commit, AI-written by the author's own account. As of 2026-10-12 nobody else had reported running it. When @djpandaman asked for a benchmark suite, @boz._. pointed here and noted it is tests, not benchmarks. Source: @djpandaman, 2026-10-12
Source: @boz._., 2026-10-12 - It assumes one CMP in the machine, or
SERIAL=...to pick one. memtest_vulkan and gpu-burn are separate projects, not included. Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09. - If
cmp-sm-testfails on every card including good ones, a different compiler changed the FP32 results; record your own baseline on a good card (bin/sm-verify record <file>, then setCMP_SM_BASELINE). Source: repo github.com/dev-boz/cmp170hx-testsuite by dev-boz on GitHub, posted by @boz._., 2026-10-09.