I'm working in a CUDA parallel algorithm and I wonder how to compute the theoretical speedup. I know that the Amdahl law isn't valid for GPUs. Does anyone know how to compute the theoretical speedup for a GPU?
Thanks
Asked
Active
Viewed 116 times
0
-
Why do you claim that Amdahl's law isn't accepted for GPGPU? There is a good answer by @aland on this point, see [Amdahl's law and GPU](http://stackoverflow.com/questions/12398929/amdahls-law-and-gpu). I think your question is based on a questionable assumption. – Vitality Jun 03 '14 at 17:42
-
Hi JackOLantern! Thanks for your reply. I've tried to compute the speedup using the equation given @aland, and I've got a speedup of 3.70. That result can't be the theoretical speedup once I've got a average speedup of 6.30 just comparing my parallel code time with the sequential time. I've got several results like this one, so I've assumed that the amdahl law isn't valid for GPUs. What do you think about it? – MSO Jun 03 '14 at 18:04