Skip to first unread message
pos...@gmail.com
Sep 11, 2014, 11:27:23 AM9/11/14
to openbla...@googlegroups.com
I compiled OpenBLAS on a recent Core i7-5930K which belongs to a new Haswell-E class of processors. I'm getting 46.5 GFlops in DGEMM (tested with N=9216) with 1 thread which is 79% of peak with Turbo frequency of 3.7 MHz.
The Haswell dgemm kernel file in OpenBLAS has 53.3 GFlops written in the comments which is 85% of peak in Turbo frequency of 3.9 GHz of your Core i7-4770. The 79% vs 85% is not that different but I'm curious what may be the reason of the poorer performance on the Haswell-E chip. Is it cache or main memory latency diferences (the system uses DDR4 vs DDR3 memory)? Maybe some of the parameters should be set differently? Which parameters should I try to change to fine tune the performance?
The Haswell dgemm kernel file in OpenBLAS has 53.3 GFlops written in the comments which is 85% of peak in Turbo frequency of 3.9 GHz of your Core i7-4770. The 79% vs 85% is not that different but I'm curious what may be the reason of the poorer performance on the Haswell-E chip. Is it cache or main memory latency diferences (the system uses DDR4 vs DDR3 memory)? Maybe some of the parameters should be set differently? Which parameters should I try to change to fine tune the performance?
Reply all
Reply to author
Forward
0 new messages