On 06/21/2017 05:29 PM, Christopher Samuel wrote: > On 21/06/17 22:39, John Hearns wrote: > >> I would speculate about single socket AMD systems, with a smaller form >> facotr motherboard, maybe with onboard Infiniband. Put a lot of these >> cards in a chassis and boot them disklessly and you get a good amoutn of >> compute power. > > I thought it interesting that the only performance info in that article > for Epyc were SpecINT and (the only mention for SpecFP was for Radeon). >
I worried about this as well, assumed AMD was bragging about Integer performance because their FP performance was worse. The opterons were somewhat weak on this front with a shared FP unit per two "cores". Turns out AMD had a pleasant surprise on that front with Epyc. I found few interesting data points: CPU chips IntRate FPRate Compiler ====== ===== ==== ====== ========== E5-2630 V4 x2 795 651 Intel compiler E5-2699A V4 x2 1800 1090 Intel compiler v16 E5-2699A V4 x2 1160 760 Gcc -O2 v6.2 Epyc 7601 x2 2100 1660 Open64 v4.5.2.1 Epyc 7601 x2 1840 1340 Gcc -O2 v6.2 Epyx 7601 x1 1200 860 Open64 v4.5.2.1 Stream numbers are quite good as well, not surprising with double the channels. While not a good indicator of overall performance, at least it's scaling the bandwidth per core so it's not a bigger bottleneck than the previous generation. Pretty impressive, AMD's fastest with gcc beats Intel's fastest with the Intel compiler. Open64 of course extends that lead. Also interesting is that the fastest Epyx single sockets do overlap with the most popular Intel dual sockets. The Epyc + Open64 numbers on on specbench.org, the gcc -O2 numbers are from AMD presented slides published on numerous tech sites after yesterday's announce. >From what I can tell of pricing it looks like if you are trying to maximize CPU2006 / node price it looks like the E5-2620 or E5-2630 on the Intel side and the EPYC 7281 on the AMD side, at least for dual sockets. It's hard to be precise though because AMD hasn't posted performance numbers for their entire line of Epyc chips yet. _______________________________________________ Beowulf mailing list, [email protected] sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
