Summing Up The AMD EPYC 7742 2P Performance In One Graphic

If you didn't have a chance since last night to check out our benchmarks of the AMD EPYC 7742 and EPYC 7502 Linux performance, I certainly encourage you to do so. Even if you aren't a server enthusiast, it's incredible to see the engineering achievement of AMD with Zen 2 and how the race is certainly back on in the CPU space. If you are short on time, here's the quick summary of our initial AMD EPYC 7002 benchmark results.

As all independent reviewers seem to be in consensus, the EPYC 7002 series is a godsend for just about everyone, sans Intel. The EPYC 7002 line-up provides the most competitive option against Intel we've seen in the server space in many years if not ever, customers have these lower-cost CPUs that actually perform better and more power efficient in a majority of real-world workloads, and Rome further benefits consumers by reigniting Intel's engine and they will certainly need to respond in some manner by either (much) better pricing or some card up their sleeves.

Better pricing seems about all Intel can do for the next quarter or two at least with Cooper Lake not coming out until next year and even then topping out at 56 cores while the next real battle could be when Icelake server parts finally ship. Intel though also has an upper-hand with a more mature and optimized software stack where they may squeeze more performance potential out of their existing chips as we've seen through their many upstream Linux contributions and the likes of Clear Linux, SVT, MKL, etc where single and even double digit percentage improvements in performance have routinely been achieved.

Anyhow, for those short on time or not yet digging through our pages of benchmark results, here is a summary of the performance between the EPYC 7742, EPYC 7601, and Xeon Platinum 8280 all in dual socket configurations and with each server running at their maximum supported memory channels and optimum frequencies. All servers freshly tested on Ubuntu 19.04 with Linux 5.2 as part of our Rome benchmarking.

As explained yesterday, the only areas where Cascadelake (green-ish line) is beating AMD's Rome is with their own MKL-DNN software (in some of the sub-tests at least), in MariaDB/MySQL strangely but that is being further investigated, and then when testing single-threaded workloads individually without fully taxing them in parallel as would be common in a production server. Cascadelake leading in the single-threaded tests isn't too surprising considering the higher clock speeds on the 8280 and then in some workloads AVX-512 does help out a lot. But as you can see, the EPYC 7742 (purple line) pretty much takes the cake. It's also striking to see how much faster Rome is than the EPYC 7601 2P.

With the dozens of tests run for this comparison, the EPYC 7742 2P has 22% better performance than the dual Platinum 8280 server based on the geometric mean or 56% compared to the 7601 2P.

And an extra graphic...

Here is also a radar chart for the performance-per-dollar graphics based on the CPUs current retail pricing. The chart here is just showing for the higher is better graphs (just a TODO item I need to do with the Phoronix Test Suite to make it show for all). But anyhow in looking at all those results, EPYC clearly wins on performance-per-dollar and even the previous-generation EPYC 7601 represents better value than the Xeon Platinum 8280 that command $10k USD a piece at retail levels.

Now see all of the EPYC 7502 and 7742 benchmarks if not already having done so while more benchmarks forthcoming. Currently wrapping up AOCC 2.0 compiler tests followed by taking Rome on a trip to FreeBSD/DragonFlyBSD, seeing how the performance compares on a few different Linux distributions, looking more at the modern CPU mitigation costs, and other benchmarking fun.