om trying to set up a small cluster for my compression runs but im worried about AMD's half count on FPU cores. (quad core only has 2 fpu cores).
is Deflate, zopfli. lzma heavy on integer or floating points calculations?
om trying to set up a small cluster for my compression runs but im worried about AMD's half count on FPU cores. (quad core only has 2 fpu cores).
is Deflate, zopfli. lzma heavy on integer or floating points calculations?
Last edited by SvenBent; 4th May 2015 at 05:29.
Generally, portable lossless compression doesn't involve floating-point numbers. But SIMD instructions can also operate on integer vectors. However, I doubt that there's significant number of compressors making use of SIMD instructions. Deflate, zopfli and lzma are LZ-codecs and they're almost fully scalar integer IMO, so the number of FPU units does not matter for them.
SvenBent (4th May 2015)
actually, intel & amd has pretty similar cores with 4 int/scalar EUs and 2 simd/fpu EUs. the greatest difference is that intel can employ all 4 int engines in the single thread while amd can't do the same. when your program is FP/SIMD-bound or when it's multithreaded and you have HT enabled, single Intel core is about the same as single AMD module. single-threaded scalar/int performance is the main thing that makes Bulldozer a mistake (omitting the fact that AMD sells you 1 core at the price of 2)