Building a small-scale GPU cluster with Parallela

[vlnguyen92]

Hello,

I'm trying to build a cluster similar to Brian Guarraci's and we're almost done. Since the Parallella boards can run OpenCL. I wonder if this kind of system could mimic the behavior of a GPU cluster? I mean a host (the Intel NUC) could schedule multiple GPU jobs to the boards.

I'm new to this and would appreciate any comments/critiques.

Thank you,

[9600]

Parallella boards don't have a GPU, so you won't be scheduling any GPU jobs to them. However, they do have an Epiphany accelerator, which is supported as an OpenCL device via COPRTHR and so you can target OpenCL kernels to this.

You could use an Intel NUC as a front-end processor / managing node for a cluster, or you could use another Parallella board, or some other host. It all depends on your application's particular requirements.

Regards,

Andrew

[vlnguyen92]

Thanks a lot for the response. Perhaps I wasn't very clear. Actually my intention is to make the Parallella boards act as GPUs (OpenCL devices as you suggested) that can mimic the behaviors of a small GPU cluster.

I haven't used COPRTHR so I'll go ahead and try that.

[9600]

Ah, interesting. Well if you want OpenCL you'll need to use the COPRTHR SDK and it offers numerous APIs, e.g STDCL also, so there may be one that is more appropriate for your particular use than OpenCL.

Regards,

Andrew