Re:Basic Blackbird Bundle with 4-core IBM Power9 CPU
by dragontamer on Tuesday November 27, 2018 @11:06AM
(#18543467)
A lot of the "weakness" there is due to the weak 4x64-bit vector units equipped on the Power9. Again: 4x64-bit vector units (Power9) simply has less throughput than 4x128-bit (EPYC) units, which has less throughput than 3x512-bit (Xeon Gold) vector units.
On the few tests which don't have anything to do with vector units, Power9 does incredibly well. See 7-Zip Compression, Stockfish, Fhourstones, LLVM compiles. Power9 is very, very far ahead of the game here.
When vector units (ie: OpenMP or intrinsics) are involved: such as Parboil, Gimp, x264, Rodinia, etc. etc. the Power9 does very poorly.
-------------
Something else I've noticed: Power9 has a lot of latency behind its units: it is always 2-cycles or 3-cycles of latency (while EPYC or Skylake have many operations with 1-cycle of latency, like add or subtract). Power9 relies upon instruction-level parallelism to get things done.
So there are a few cases like the Rust Mandelbrot benchmark, which surely hits that latency problem, while the Rustlang Prime benchmark shows off the ILP goodness the system offers.
On the few tests which don't have anything to do with vector units, Power9 does incredibly well. See 7-Zip Compression, Stockfish, Fhourstones, LLVM compiles. Power9 is very, very far ahead of the game here.
When vector units (ie: OpenMP or intrinsics) are involved: such as Parboil, Gimp, x264, Rodinia, etc. etc. the Power9 does very poorly.
-------------
Something else I've noticed: Power9 has a lot of latency behind its units: it is always 2-cycles or 3-cycles of latency (while EPYC or Skylake have many operations with 1-cycle of latency, like add or subtract). Power9 relies upon instruction-level parallelism to get things done.
So there are a few cases like the Rust Mandelbrot benchmark, which surely hits that latency problem, while the Rustlang Prime benchmark shows off the ILP goodness the system offers.