Cascade Lake vs. Rome With MrBayes, dav1d 0.5, OSPray, SVT-VP9, OIDn + Other Benchmarks

While swapping around CPUs for the AMD EPYC vs. Intel Xeon Cascade Lake testing of Facebook's RocksDB enterprise workload testing, I also took the opportunity for running some other recently updated test profiles on these EPYC/Xeon parts under test.

These newest results shouldn't be particularly surprising but are primarily just benchmark results for some updated versions of existing tests. With recently a number of updated test profiles on OpenBenchmarking.org against the upstream programs under test, here are simply those latest performance numbers when running on the dual Intel Xeon Platinum 8280 and the AMD EPYC 7601 Naples and EPYC 7502 / 7642 / 7742 Rome processors all in 2P configurations. The setups were the same as from the RocksDB testing with running the newly-minted Ubuntu 19.10.

With the MPI-based MrBayes scientific software updated against v3.2.7 and focused on Bayesian inference of phylogeny, the Rome CPUs were ahead of the Xeon Platinum 8280. That's even with MrBayes supporting AVX-512 optimizations. The fastest processor in this test was the EPYC 7502 due to its higher base clock frequencies and MrBayes benefiting from that as opposed to the extra cores. There's a nice boost in performance here for this popular molecular biology workload in going from Naples to Rome.




Dav1d 0.5 was released last week as the newest version of this AV1 video decoder. Even with more AVX tuning and other performance work going into dav1d 0.5, the Rome CPUs still delivered the best performance by far.








We've been benchmarking Intel's OSPRay renderer more with that becoming part of the Intel oneAPI toolkit and that beta expected this quarter. Most often in OSPray the Xeon Platinum 8280 performance was comparable to the EPYC 7502 CPUs. The only exception -- and where the Xeon Platinum performance was leading -- was when testing the "magnetic reconnection" scene that seems to do extremely well on Intel CPUs possibly due to hitting some AVX-512 fast paths well or so.

Likewise, the Xeon Platinum performance did lead when running Intel Open Image Denoise, but not surprising.



Intel's software team earlier this month made their first tagged release of the SVT-VP9 video encoder. Similar to seeing Ryzen 3000 series leading on the desktop CPU front, EPYC CPUs were winning over the Xeon Platinum 8280 for this well-threaded video encoder. The EPYC 7502 was doing the best of the tested CPUs with its higher base clock speeds for the sample content used. Besides the EPYC 7002 series coming out in front of Xeon Platinum, they continue showing significant improvements compared to the EPYC 7601 Naples.

With the ACES DGEMM version out of Los Alamos National Lab, the Xeon Platinum 8280 could only tango with an EPYC 7642.

With our popular "timed Linux kernel compilation test" we have upgraded that test profile to tracking Linux 5.4 development due to the upstream state of that building nicely now with LLVM Clang 9.0+ without any extra patches. While this test was using Ubuntu 19.10's default GCC 9 compiler, it's great seeing the EPYC 7742 2P still being able to build the mainline kernel in under 20 seconds! Previously the test profile was tracking the Linux 4.18 LTS kernel and even with all the new code from 4.18 to 5.4, these build times are still incredibly short.


If you want to compare your own Linux system(s) performance to these extra benchmarks found in this article, simply install the Phoronix Test Suite and run phoronix-test-suite benchmark 1910176-HU-EXTRA098313 for your own fully-automated, side-by-side benchmark comparison.