A colleague from our team noticed that, under Ubuntu Linux, for several graphics card models it appears problematic to autodetect the correct RAM hardware which results in the exclusion of these machines from computation in GPUGRID.

One example: Instead of reporting 2 GB, for a GTX 640 card Linux will only recognize 199x MB - no taks is received from the GPUGRIUD server. Using the same card in a Windows system results in correct recognition of 2048 MB - taks are received.

I was wondering whether it would be possible to change the task delivery criteria on the GPUGRID server from 2 GB of RAM to 1990 GB? Additional compute power could be acquired for the project...

Michael.

P.S.: Curiously, even Windows 10 reports only 4096 MB of GPU RAM for my 6 GB GTX 1060 Graphics card...

---

President of Rechenkraft.net - Germany's first and largest distributed computing organization.

I believe there are several ways the Work Server may ban a card. E.g.,

<min_nvidia_compcap>MMmm</min_nvidia_compcap>
minimum compute capability

https://boinc.berkeley.edu/trac/wiki/AppPlanSpec

The GTX 640 has a compute capability of 2.1 (GDDR3) or 3.5 (GDDR5).

Click on "CUDA-Enabled GeForce Products": https://developer.nvidia.com/cuda-gpus

https://en.wikipedia.org/wiki/CUDA

As described above, the memory detection under Ubuntu Linux remains incorrect - even in graphics card models with up to date compute models.

Michael.

---

President of Rechenkraft.net - Germany's first and largest distributed computing organization.