Re: max number of CPU PEs per GPU?

From: Jim Phillips (jim_at_ks.uiuc.edu)
Date: Fri Jun 10 2011 - 15:15:19 CDT

You're probably hitting some driver limit on the number of client
processes that can connect to a single GPU. With CUDA 4.0 we'll be moving
to a multi-threaded model that will share GPUs more efficiently that
should eliminate this issue. Even then 48 Opteron cores might not get any
benefit from a single GPU.

-Jim

On Thu, 9 Jun 2011, Danny Xu wrote:

> Hi,
>
> I was testing the namd 2.8b2 cuda binary (Linux-x86_64-CUDA) on a 48-core
> opteron/1x GTX480 server. I got:
> ...
> Pe 41 sharing CUDA device 0 first 0 next 42
> Pe 40 sharing CUDA device 0 first 0 next 41
> FATAL ERROR: CUDA error cudaStreamCreate on Pe 23 (chili device 0): no
> CUDA-capable device is available
> ------------- Processor 23 Exiting: Called CmiAbort ------------
> Reason: FATAL ERROR: CUDA error cudaStreamCreate on Pe 23 (chili device 0): no
> CUDA-capable device is available
>
> First I thought there was something wrong with the GTX480 card. Then I tried 12
> PE's and the job ran just fine. It turned out the max number core I could use
> with one GTX480 is 14. My question is, is this hard coded in namd 2.8? Or it's a
> bug in 2.8b2 and has been fixed in 2.8?
>
> It'd be nice if all 48-core could be used with the GPU.
>
> Thanks,
>
> DX

This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:20:25 CST