Slow mem copy speed

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

SizeDataForGrid = 36; 

unsigned int hTimer1; 

CUT_SAFE_CALL( cutCreateTimer(&hTimer1) ); 

cutResetTimer(hTimer1); 

cutStartTimer(hTimer1); 

CUDA_SAFE_CALL( cudaMemcpy( grid[0].blocks, 

(device_DataForGrid+0*SizeDataForGrid), SizeDataForGrid, 

cudaMemcpyDeviceToHost) ); 

cutStopTimer(hTimer1); 

printf("\nMemory copy time(for tracking. From device): %f msec", 

cutGetTimerValue(hTimer1));

This code on host side, executing 36msec.

Why it very slow?

2

Have you tried bandwidthTest CUDA SDK sample on your system? With small data sizes memcpy speed (in all the three cases) is defined by OS/chipset/driver to a great extent.

3

bandwidthTest - showing results in Mbit/sec or in Mbyte/sec ?