Amazon Graviton3 vs. Intel Xeon vs. AMD EPYC Performance

First up was the HPCG benchmark from Sandia National Labs. Even Graviton2 in this HPC benchmark came out far ahead of the similarly sized Xeon and EPYC instances while with Graviton3 it is 33% faster than Graviton2, or 3x the speed of the Intel Xeon Ice Lake 4xlarge instance. It is important to remember though with the Graviton-based instances each vCPU is backed by a physical CPU core rather than SMT/HT with AMD and Intel.

Not only does Graviton have the advantage of being all physical CPU cores without SMT, but Amazon's in-house processor can lead to a hefty price savings in the hourly on-demand rate and the performance-per-dollar.

In other HPC workloads meanwhile the Graviton3 advantage tended to not be to an extreme as seen by HPCG. With NASA's NPB tests, the Xeon instance tended to lead while the EPYC Milan instance was putting up a good fight as well.

In the NPB test cases, the Xeon Ice Lake instance was able to prove itself to be of good value too for this set of tested workloads.

For the LC0 neural network chess engine, the Intel Xeon Ice Lake was the fastest of the tested 4xlarge instances -- with this workload able to leverage AVX-512. Meanwhile the Graviton3 was now performing inline with the AMD EPYC Milan instance under LC0 as quite a feat and exhibiting great performance uplift from Graviton2.