uqsm5090

Fork of the scitix uqsm v1.1 stress-test harness with one change: the GPU module can run with gpu-burn instead of dcgmi diag -r 4, so it works on RTX 5090 / Blackwell hosts where DCGM does not yet support the SKU.

Getting it — use the OSS tarball (recommended), not git clone

There are two sources and they are not equivalent:

Source	gpu-burn CUDA libs included?	Ready to run?
OSS tarball (recommended)	✅ yes — gpu/lib/*.so.12 are bundled	Unzip and run, no extra step
git clone (developers)	❌ no — libcublasLt.so.12 is ~717 MB, over GitHub's 100 MB/file limit	must run gpu/setup_libs.sh once to fetch the libs

Recommended (unzip-and-run):

wget https://oss-cn-shanghai.siflow.cn/scitix-release/uqsm5090_v1.1.tar.gz   # Shanghai
#   or: https://oss-ap-southeast.scitix.ai/scitix-release/uqsm5090_v1.1.tar.gz  # Malaysia mirror
tar xzf uqsm5090_v1.1.tar.gz && cd uqsm5090
./gpu/run_sm.bash gpuburn 60        # runs immediately — libs are already in gpu/lib/

From git clone (only if you are modifying the code):

git clone https://github.com/404mario/uqsm5090 && cd uqsm5090
./gpu/setup_libs.sh                 # one-time: downloads the cuBLAS libs (~828 MB) that can't live in git
./gpu/run_sm.bash gpuburn 60

The big cuBLAS .so files physically cannot be committed to GitHub (single-file 100 MB limit), which is why a clone needs setup_libs.sh but the OSS tarball does not. If you just want to run the test, use the OSS tarball.

Hit (DIED!), or unsure which build matches your driver? See docs/COMPATIBILITY.md — the driver ⇄ CUDA ⇄ artifact matching matrix and what the runtime preflight checks. Background: docs/POSTMORTEM-rtx5090-gpuburn.md.

Modules

• cpu/ stress-ng --matrix 0 -t 10m (unchanged)
• mem/ memtester via memtester_loop.sh 88 32 65536 (unchanged, ~7h)
• ib/ ubimonitor -t -q 10 (unchanged)
• gpu/ either dcgmi diag -r 4 or gpu_burn -d <sec>

Usage

./uqsm.bash                                    # all modules, auto GPU tool
./uqsm.bash -dt gpu                            # GPU only, auto tool
./uqsm.bash -dt gpu --gpu-tool gpuburn         # force gpu-burn
./uqsm.bash -dt gpu --gpu-tool dcgm            # force dcgmi diag -r 4
./uqsm.bash --gpu-tool gpuburn --gpu-duration 1800

Auto-detect rule: if nvidia-smi -L reports any RTX 50xx / Blackwell card the default is gpuburn; otherwise it is dcgm. Override with --gpu-tool.

Results are written to result/<module>_result.json and rolled up into result/report.json.

Building gpu-burn for sm_120 (RTX 5090) — no Docker, no nvcc

The gpu_burn binary and compare.fatbin must live in gpu/. The build wrapper needs only g++, python3, and PyPI access — no Docker, no CUDA toolkit, no nvcc, no sudo:

~/uqsm5090/gpu/build_gpuburn.sh

It pulls the CUDA 12.8 headers, cuBLAS .so, and libnvrtc from NVIDIA's pip wheels, compiles the g++ driver, links against libcublas.so.12, and regenerates gpu/compare.fatbin from compare.cu as sm_120 SASS via nvrtc (no prebuilt kernel is reused — a fatbin is toolchain-version-locked, and a CUDA-13 one fails on a 570 driver with CUDA_ERROR_UNSUPPORTED_PTX_VERSION).

Why CUDA 12.8 and not the newest toolkit? 12.8 is the first toolkit that supports Blackwell sm_120 and only requires NVIDIA driver ≥ 570, yet it also runs on every newer driver — so a single 12.8 build is portable across the whole RTX 5090 fleet. A CUDA 13 build needs driver ≥ 580 and otherwise kills every gpu_burn worker ((DIED!)) on a 570-era host. run_sm.bash runs a driver preflight and the scripts are CUDA-major-agnostic (override CUDA_VER / CUDA_MAJOR to target a different driver); see CLAUDE.md for details.

Running gpu-burn (bare metal, no Docker)

gpu-burn runs directly on the physical machine — Docker is not used at runtime. The only CUDA toolkit dependency is libcublas.so.12 (which pulls in libcublasLt.so.12); both are vendored under gpu/lib/ and loaded via LD_LIBRARY_PATH. The CUDA driver lib libcuda.so.1 comes from the installed NVIDIA driver — it is the user-space half of the kernel module and can never be bundled or statically linked.

./uqsm.bash -dt gpu --gpu-tool gpuburn --gpu-duration 600   # no sudo needed

The libs (gpu/lib/libcublas.so.12 + libcublasLt.so.12, ~828 MB) are too large for GitHub's 100 MB limit, so they are not committed. They are already bundled inside the OSS release tarball (below). After a bare git clone, fetch them once — no Docker, no CUDA toolkit, just curl:

./gpu/setup_libs.sh        # downloads the cuBLAS pip wheel and extracts the two .so

gpu/build_gpuburn.sh is only needed to rebuild the gpu_burn binary itself; the prebuilt binary is committed, so normal use never rebuilds anything.

Using the dcgm backend (non-Blackwell hosts)

The DCGM .deb for CUDA 13 is 419 MB and is not committed to git (exceeds GitHub's 100 MB single-file limit). Download manually and drop into gpu/ before running with --gpu-tool dcgm:

# DCGM 4.5.2 packages for Ubuntu 22.04 / amd64
cd ~/uqsm5090/gpu
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2204/x86_64/datacenter-gpu-manager-4-cuda13_4.5.2-1_amd64.deb
# (core .deb is small and already committed)

Or use any DCGM build matching your CUDA major version; gpu/run_sm.bash globs datacenter-gpu-manager-4-*_amd64.deb.

Smoke test result (8x RTX 5090, 600 s, gpu-burn)

Tested 8 GPUs:
    GPU 0..7: OK
===GPU Stress Test Success===