I mentioned CUDA, but this might also apply to ATI boards.

I have a pair of GTX1070 that have 8gb of memory. Looking at the event messages, it is clear that OpenCL only uses just under 4gb of memory. I assume that is a restriction based on 32 bit addressing. That is just a guess.

It appears I can run 4 Einstein or 5 milkyway on any GPU that has 4gb of memory. Less concurrency for 3 and 2gb graphics boards. I read that miners have specific programs they run to mine on the larger gtx1080 that have 11gb of memory and they use all of the 11gb. I was wondering if any BOINC projects are coded to use all available GPU memory and not be restricted to 4gb. I assume that would be more efficient but perhaps a larger burden (heat) on the cooler.

I tried running more than 4 tasks on my gtx1070 and the tasks started erroring out. I assume the problem is the 4gb OpenCL. My systems are Core2 quad so I only have 4 cores. Maybe they errored out because they were not fed properly by the cpus and not the 4gb OpenCL limit?

Actually, OpenCL can only use 1/4 of the gpu onboard memory. The limitation is in the OpenCL framework and not specific to any projects OpenCL applications.

GPUGrid is CUDA only. They have limited tasks and finally updated their linux app today that should work.

You asked about AMD, well... Moo! still has CAL capable work available that AMD cards can typically run. I have a few really old non-OpenCL capable devices here and there that sometimes runs it.

---