Benchmarking Amazon's Graviton2 Performance With 64 Neoverse N1 Cores Against Intel Xeon, AMD EPYC
When comparing Amazon EC2 between M6g, M5, and M5a general purpose instances, the new Graviton2-based performance exceeded my expectations especially with being not impressed by the original Graviton processors and their A72 cores, but Graviton2 with up to 64 Neoverse N1 cores delivered quite the performance uplift. The M6g instances generally offered comparable or better performance to the x86_64 general purpose current-generation instances, but note that in those configurations EC2 is offering a combination of physical+SMT vCPUs compared to all physical cores with M6g in coming to the same vCPU count. In any case, for the same instance size, the M6g was a surprisingly strong contender for most workloads. Unfortunately due to EC2 costs, the comparison was limited in nature compared to the Phoronix Test Suite offering hundreds of possible tests to run.
When taking the geometric mean of all the tests run on the various general purpose instances, the Graviton2 instances came out well ahead of the similarly sized M5 Intel and M5a AMD instances. For the 2xlarge/4xlarge also supported by the original Graviton, the Graviton2 performance was about twice that of the previous generation.
Meanwhile when it came to evaluating the bare metal performance of Graviton2, Amazon's Graviton2 with sixty-four Neoverse N1 cores was no match to the top-end Intel Xeon and AMD EPYC processors. With the bare metal performance, the EPYC 7002 series generally swept up all the wins with up to 64 cores / 128 threads per socket. In a number of workloads though the Graviton2 bare metal server was at least offering some competition to these current Intel Xeon Scalable and AMD EPYC Rome offerings and much better than we previously saw out of the original Graviton 2 CPUs.
Granted, with Amazon not expressing any plans to offer the Graviton2 SoC outside of their cloud, the overall positioning of the Graviton2 CPU isn't as important as the instance comparison.
That's it for the Graviton2 benchmarking for now. If you enjoy these tests, consider joining Phoronix Premium to help support the testing operations or at the very least to disable any ad-blocker when viewing this site. If/when funds allow, an even larger Graviton2 comparison with more workloads will be forthcoming.
If you enjoyed this article consider joining Phoronix Premium to view this site ad-free, multi-page articles on a single page, and other benefits. PayPal or Stripe tips are also graciously accepted. Thanks for your support.