MPI and CUDA

[Windoze]

Prof. Creel mentioned CUDA here¹⁾ a while back.

The Nvidia forum has a recent thread that might be of interest:

Mixed Programming, Combining MPI and CUDA


MPI needs no explanation here, but CUDA might.

CUDA is Nvidia's development platform for their inexpensive, multi-processor, multi-register GPU's. These devices, found in video cards from many manufacturers, can be used as serious number crunchers. My card for example, a GT 9800, has 14 multiprocs and over 8,000, thirty-two bit registers. It can handle more than 12,000 simultaneous threads. And it's just a cheaper card (USD 120).

On my dual-core Intel, running eight simultaneous, near real-time FFT's, or ten repeated 4k x 4k matix, 3-vector multiples barely made the main board's cpu fan speed up. Although the boxes' total power consumption went from 250 to 320 Watts :)

cheers,
- Howard in Florida
___________________
¹⁾ More then 50GFlops

[Windoze]

Can't help drooling here

"Supercharge Your Research" - Nvidida press release

"Tesla Preconfigured Clusters consist of x86 CPU servers coupled with Tesla S1070 1U GPU systems. Configurations start at 16 Teraflops of performance, delivered by four Tesla S1070s, each containing four Tesla 10-series GPUs. All systems include host servers, infiniband switches, cabling and are fully customizable to suit individual customer needs"