CUDA Transfer speeds and Hypertransport

Accelerated Computing CUDA CUDA Programming and Performance
1

Please find in the attachment the correspondence with NVIDIA support about the above subject.

Any ideas / opinions from the group whether CUDA transfer speeds will improve with mGPUs with hypertransport technology ?

Thanks
hmovva

[attachment=6300:attachment]
hypertransport.txt (3.77 KB)

2

Even with ultra-fast HT link speeds, you are still moving data from the GPU to the system RAM: that RAM will then become the bottleneck (this is already seen in PCIe gen2 boards paired w/ gen2 MB’s, search for the benchmarks on the forums).

3

Thanks for the reply. I assume mean to say system DRAM is the bottleneck, because device to device transfer rates exceed 40GB/sec. In that case,whether use of DDR3 for sytem will help as and when they become available? Thanks again.