Linked by Thom Holwerda on Fri 7th Jul 2006 13:22 UTC
Thread beginning with comment 141165
To view parent comment, click here.
To read all comments associated with this story, please click here.
To view parent comment, click here.
To read all comments associated with this story, please click here.





Member since:
2006-01-21
Most of that improvement comes from ICC being a simpler compiler (it does C/C++ to x86/x86-64) compared to GCC, which handles far more languages and architectures and pays a performance penalty for it.
Without getting into a flame war over licenses, I'll just point out that Intel's icc and Sun's cc are good optimizing compilers; by comparison, gcc is mediocre. You get what you pay for.
All that said ... autovectorization only helps in benchmarking. Any real world code either has so little vectorized code that there is no difference, or is already hand-tuned assembly, or has to be sufficiently cross-processor-generation that processor-specific optimizations like autovectorization are turned off. The market for benchmark-inflating but not useful optimizations isn't worth AMD's time.
(A supercomputer is going to use BLAS or some other optimized math library which already has any relevant vectorization; AMD has BLAS for their processors.)