Hi, I run the bandwidth test provided in the CUDA toolkit on my Tesla C1060 card. The value obtained is of 73259 MB/s. Is this the “effectvive bandwidth” value (the theoretical value is around 102 GB/s)? Many thanks, jony
bandwidthTest uses
cudaMemcpy( d_odata, d_idata, memSize, cudaMemcpyDeviceToDevice)
to test device-device transfer.
It includes read and write, so bandwidth is about 70~80 % of theoretical bandwidth.
you can write a kernel to copy a data, then you will find that you cannot achieve full bandwidth.
However if you only read data, then you can reach higher bandwidth.