3 ms·

Benchmarks sound like their fighting hardware frequency throttling more then actually benchmarking the chips. One big issue to point out: >Floating Point: NAM

Benchmarks sound like their fighting hardware frequency throttling more then actually benchmarking the chips. One big issue to point out: >Floating Point: NAMD GCC sucks at automatic hardware vectorization. So does the LLVM. Really the only time you can count on getting automatic hardware vectorization is if you shell out for the ICC AND write your code in Fortran. I'm gonna bet they didn't vectorized a goddamn thing, but we can't inspect anand's binary so we'll never know. The results are still likely correct, but IBM should have lost by a smaller margin. TL;DR POWER8 is fun but costs 5k more then Xeon per rack mount and uses about 2x the power usage for 10% less performance on generalized work loads. But can pull off 10-15% more performance on _some specialized_ workloads. So meh?

vonmoltke 11y ago
My experience has been that icc (the C compiler) does a very good job of auto-vectorization. xlc does as well, provided you avoid the horribly misdesigned POWER6 architecture. I agree that gcc sucks for this; we didn't have LLVM last time I was doing this kind of work, so I can't comment on that.
- valarauca1 11y ago
  LLVM is about on par with the GCC. You just keep moving variables around it decides the moons of Jupiter are properly positioned to vectorize your code. System intrinsics are always your best bet, just really annoying.
  - vonmoltke 11y ago
    We ended up doing some hand-rolled vectorization that way on the POWER7, because SIMD only really worked well if the data was aligned with the SIMD register width. Intel fixed that problem with a relatively low-cost unaligned load instruction.
- CyberDildonics 11y ago
  If you want to really use SIMD units (on x86) check out ISPC. I recommend writing small functions in it that work over large chunks of memory so that SIMD can run at full speed with good cache locality. This ends up being -very- fast.
  - valarauca1 11y ago
    Link for the lazy: https://github.com/ispc/ispc/ https://github.com/ispc/ispc/
spooningtamarin 11y ago
#pragma omp simd collapse(N) works for me almost every time.

vonmoltke 11y ago

My experience has been that icc (the C compiler) does a very good job of auto-vectorization. xlc does as well, provided you avoid the horribly misdesigned POWER6 architecture. I agree that gcc sucks for this; we didn't have LLVM last time I was doing this kind of work, so I can't comment on that.

valarauca1 11y ago
LLVM is about on par with the GCC. You just keep moving variables around it decides the moons of Jupiter are properly positioned to vectorize your code. System intrinsics are always your best bet, just really annoying.
- vonmoltke 11y ago
  We ended up doing some hand-rolled vectorization that way on the POWER7, because SIMD only really worked well if the data was aligned with the SIMD register width. Intel fixed that problem with a relatively low-cost unaligned load instruction.
CyberDildonics 11y ago
If you want to really use SIMD units (on x86) check out ISPC. I recommend writing small functions in it that work over large chunks of memory so that SIMD can run at full speed with good cache locality. This ends up being -very- fast.
- valarauca1 11y ago
  Link for the lazy: https://github.com/ispc/ispc/ https://github.com/ispc/ispc/

spooningtamarin 11y ago

#pragma omp simd collapse(N) works for me almost every time.