6

OS is Debian. Looking for a quick command line tool to drive GPU load to max to test temps / cooling.

I already looked at nvbench, which does compile and run, but seems more like getting useful benchmark numbers vs maxing out the GPU. (util only hit 50%)

I'm just looking for something simple that can max out my gpu. Even just "here's some cuda code that will max your gpu, compile like this with nvcc, etc".

cat pants
  • 299

3 Answers3

4

You could try gpu-burn :

Multi-GPU CUDA stress test http://wili.cc/blog/gpu-burn.html

The project site describes how to use it with Docker, but you could use it simply like this :

git clone https://github.com/wilicc/gpu-burn
cd gpu-burn
make
./gpu_burn 300      # run gpu_burn for 300 seconds
harrymc
  • 498,455
1

Adding to the ways to stress test it, don't forget to test the memory, the VRAM. Actually, that might be the first test you want to run. Clearing the memory as functional is important if you intend to rely on the card, and are going to stress test the GPU itself.

https://github.com/GpuZelenograd/memtest_vulkan

See also https://simtk.org/projects/memtest https://sourceforge.net/projects/cudagpumemtest/

This type of program I've used, but it's less valuable as it's only testing a subset of the vram, whatever happens to be available.

https://github.com/ihaque/memtestCL

Because of that I tend to use CPU, GPU vram and GPU test or exercise routines from booted ISOs rather than any booted OS, but the MemtestCL is probably sufficient if you're not running a repair center and certifying cards for a manufacturer or re-manufacturing them.

https://angryadmin.sesc.dev/posts/nvidia-vram/

This shows some Nvidia specific tools are available and run on a thin linux base rather than DOS, they test some logic, not only eg. Performance or Vram integrity.

extract: "MATS stands for memory automated test system and MODS stands for modular diagnostics software"

Xenek Stoehr
  • 49
0

Here's another suggestion:

Try a crypto miner like ethminer. Works with OpenCL or CUDA (v8 - v10, not sure what happens with v11). Download it here, under Releases. Find usage examples under "docs":

.\ethminer -P stratum1+tcp://ETH_WALLET.WORKERNAME@eth-eu1.nanopool.org:9999

Depending on the mining pool you connect to, it might ask you to input an actual ethereum address (42 characters, hexadecimal: 0xbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb)

The algorithm used for mining ethereum - and most other algorithms for mining crypto currencies - are memory intensive, you can notice that by looking at your GPU memory usage (should be around 100%). Core usage will be much lower. So this is mostly useful for stress testing GPU memory and memory temps.

GChuf
  • 1,327