His points may be valid (that some people cut corners to get the best numbers for their papers); however, as someone who spends a lot of time writing GPGPU code

His points may be valid (that some people cut corners to get the best numbers for their papers); however, as someone who spends a lot of time writing GPGPU code[1], I'll point out that MATLAB isn't the best way to benchmark GPU-based code. I'm not knocking MATLAB here, but simply pointing out that MATLAB is going to have overhead compared to, say, CUDA. Given that most GPGPU kernels run pretty quickly (less than a second, and probably under 100ms) even a little bit of overhead is going to severely affect your results. [1] See my profile.