http://dl.acm.org/citation.cfm?id=1816021 is also an enlightening paper on optimized CPU vs GPU code. It debunks the 'GPU provides 100x speedup over CPU' myth as these comparisons are using an unoptimized CPU code (without vector intrinsics, etc) against highly optimized CUDA code. It also explores various benchmarks of common scientific computations in which GPU speedup is less than 4X.
The key takeaway from the 'Considerations When Evaluating Microprocessor Platforms' is that it is difficult to accurately benchmark performances of the same code running on different platforms due to the architectural differences (which affect data/compute bounds).
In Considerations When Evaluating Microprocessor Platforms, the authors look at how to appropriately compare workload / microprocessor performance across workloads and architectures.
http://dl.acm.org/citation.cfm?id=1816021 is also an enlightening paper on optimized CPU vs GPU code. It debunks the 'GPU provides 100x speedup over CPU' myth as these comparisons are using an unoptimized CPU code (without vector intrinsics, etc) against highly optimized CUDA code. It also explores various benchmarks of common scientific computations in which GPU speedup is less than 4X.
The key takeaway from the 'Considerations When Evaluating Microprocessor Platforms' is that it is difficult to accurately benchmark performances of the same code running on different platforms due to the architectural differences (which affect data/compute bounds).