I was trying to measure the memory bandwidth of my GPU, and I used the bandwidthTest in CUDA samples.
However, after the testing, it says “The CUDA Samples are not meant for performance measurements. Results may vary when GPU Boost is enabled.” So can we use the CUDA samples to measure the performance? If not, what’s the best way of measuring the memory bandwidth (GPU->GPU, GPU->CPU, CPU->GPU)?
I use the bandwidthTest regularly as a performance measurement (an estimate of available device memory bandwidth).
Since the area of performance measurement is a complex topic and there may be many different viewpoints about what is “correct”, the disclaimer is a general acknowledgement that there may be any number of issues with a particular sample code.
Most sample codes are intended to demonstrate particular methodologies, and are not necessarily optimized to give maximal performance in a particular scenario.