cb95014 wrote:
Thank you! One of the things that I really like about the SPCR forums is the decorum. If I want to read flames, there are plenty of other sites.
Well, it that case I assume my tone was inappropriate and apologize to the other readers inconvenienced. I guess I got carried away a bit.
I'll just leave the argument "as is", everything necessary has been said, anyway.
Now, with that out of the way, back to the benchmark/load generator issue:
To recapitulate, the Intel Linpack benchmark provides a heavier load than prime95 on both AMD and Intel CPU's. However, the benchmark originates from Intel's MKL library, which in my experience [I'm mostly using the eigenvalue routines] is heavily optimized for Intel's CPUs and provides suboptimal
performance on AMD CPU's.
At least to me, that leaves unclear whether the
power figures obtained with this tool are reliable.
I would offer to code up a similar benchmark to the Intel Linpack program (solve a NxN dimensional linear equation system, several passes, etc.). This would be compiled once as a "platform agnostic" reference implementation, using the netlib reference BLAS and LAPACK implementation, and gfortran 4.4 -mtune=generic.
Then I would provide binaries optimized for each platform, which means compiled with ifort/*a-certain-non-free-for-educational-use-compiler*, FLAME and the Intel/AMD version of the Goto-BLAS. These are the compiler/library combinations that deliver the best performance (yes, better than MKL), at least for the eigenvalue problems I usually deal with.
However, I do not have (private) access to neither recent Intel CPUs nor a power meter, and do not intend to buy either. So I would need someone that could actually run the benchmarks and report results. That would obviously include running obscured binaries from me - while I will naturally provide the source and Makefile, the person would need to jump through quite some setup steps to actually build the benchmark. I just happen to have done that from my last library and compiler hunting session and it is not exactly what I would call "fun"
But before I start writing that, I would like to see someone provide power figures for prime95 and the Intel Linpack benchmark on both AMD and Intel CPU's. If the differences there seem consistent then I guess I could save the effort (and would suggest to consider using the Linpack benchmark in the future).