New record for CUDA usage in GIS

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everybody,

This went out on the wire Monday, so I thought folks here might be interested in it…

[url=“http://www.manifold.net/info/pr_gpu_record.shtml”]http://www.manifold.net/info/pr_gpu_record.shtml[/url]

I know that multi-GPU usage is routine for many of us, but using this many processors in a commercial GIS application is genuinely a new thing. Also interesting is the heterogeneous parallelism in play, not just utilizing many threads in the CPU to dispatch into the GPUs but more interestingly that a job can be dispatched into multiple different types of GPU (say, a mixed system with both G100 and G200 series GPUs). Anticipating in advance the next question… the machine was running Windows XP x64.

I hope this helps anyone who needs to convince their management that the time is NOW to make a serious investment in leveraging CUDA. [Well… anyone except our competitors, that is! :-) ]

2

Nice announcement! There is a typo in the “Why is the hardware used so inexpensive?” answer. You refer to the GTX 285 as having 480 stream processors when you mean the GTX 295.

3

Very nice… thanks for the info…

I was wondering whether you can shed some light on what I’ve asked in a previous thread: http://forums.nvidia.com/index.php?showtopic=91912

regarding the amount of CPU cores vs the GPUs on the system.

I see that you’ve launched more CPU threads then CPU cores and wondered if you didnt see performance degragation due to this configuration?

I still didnt get the machine with 4 295 with 1 dual CPU to test, but I tested 2 295 with a quad and with a dual CPU… I did see some 20-30% performance lost

when using the dual CPU (i.e. two GPUs working on one CPU core)

thanks

eyal

4

Thanks for the note! The typo is fixed now.

5

[attachment=11482:threads.png]

I visited the thread and I don’t have anything to add to the comment that the number of CPU cores does not control the number of GPUs you can use.

No.