Please,
When my RTX 5070 calculates Axiom work units (WUs), its workload is 0%, and each WU is calculated in about 15 seconds (for a reward of 1 point per WU).
So I tried increasing the number of WUs calculated in parallel on my GPU.
And I'm seeing very strange behavior, never before observed in a BOINC project : on this "small" GPU, I can even run 64 WUs in parallel, or even 100. Calculation time for each WU : about 65 seconds. GPU load : still 0% (still for a reward of 1 point per WU).
Are you absolutely certain there isn't a setting to adjust for your GPU WUs to use them (much) more efficiently ?

Copied this from one of your output files:

[GPU] This GPU binary requires NVIDIA driver 525+ for CUDA 12.x

Don't know it helps.

On my RTX 5070 I'm getting up to 750 credits depending on the experiment being run.

The GPU apps have gone back and forth with utilization like you see to 100% GPU load and back. I had done the same as you and ran multiple tasks at once on the GPU as it was only loading the CPU with a split second of GPU usage. That was many versions ago though.

I'm correcting my previous message : running 64 WUs simultaneously on the GPU only works with WUs "exp_schnakenberg_crossed_feed_diffusion_gpu".
This does not work with WUs "exp_tailspike_covariance_localization_gpu", with which I can only do about ten in parallel.

In reply to WTBroughton's message of 10 Mar 2026:

Copied this from one of your output files:

[GPU] This GPU binary requires NVIDIA driver 525+ for CUDA 12.x

Don't know it helps.

On my RTX 5070 I'm getting up to 750 credits depending on the experiment being run.

My driver version : 570.211.01, CUDA Version: 12.8
So it can't be coming from that.
Thanks !

This is a known issue with newer GPUs running CUDA 12.8 (like your RTX 5070). The GPU binary your host had was built with CUDA 12.4 NVRTC builtins, but your system's NVRTC 12.8 looks for libnvrtc-builtins.so.12.8 which wasn't present. This caused every GPU kernel to fail to compile, so the experiments fell back to error handling and finished in seconds with minimal credit -- hence the 0% GPU load.

We've patched a fix (v6.25) that properly symlinks the system's NVRTC builtins so CuPy can find them. Your host should automatically download the new version on its next scheduler contact. My server will auto-update me with your GPU results as soon as the first one comes in so I can verify everything is working correctly.

Many thanks for your answer !
I'm waiting to see if the situation improves.