Amazon's New EC2 M7a AMD EPYC "Genoa" Instances Deliver Leading Performance In The AWS Cloud

[phoronix]

Phoronix: Amazon's New EC2 M7a AMD EPYC "Genoa" Instances Deliver Leading Performance In The AWS Cloud

While back in November was when AWS originally announced new EC2 instances powered by 4th Gen AMD EPYC "Genoa" processors, only this week did they bring their M7a general purpose instances to a general availability state where anyone can access them. Being very impressed with 4th Gen EPYC bare metal as well as with Azure's HPC cloud, I fired up some benchmarks of the new Genoa-powered EC2 M7a instance compared to the new M7i instances powered by Intel Xeon Scalable "Sapphire Rapids" as well as showing how the competition is to Amazon's in-house Graviton ARM-based server processors.

https://www.phoronix.com/review/aws-m7a-ec2-benchmarks

[manithree]

Wow! Great benchmark with great info. I'm collecting some articles to convince my employer to pay for a subscription. This one just hit the top of that list. Great work.

[partcyborg]

Fantastic article!

Looking forward to putting our db product at work through some tpch benchmarks on the new generation now that both AMD and Intel instances are available. Up through the previous generation Intel has been faster core for core, but I have hopes that this is the generation that I can prioritize AMD instance scheduling over Intel

[yump]

Since the article mentioned that m7a doesn't use SMT anymore, I was wondering how these match up on physical core count. Going by this, assuming the defaults were used, c6a (Zen 3) and m7i (Sapphire rapids) would have SMT, and everything else does not, which would explain the unexpectedly large difference between Zen 3 and Zen 4.

I'm surprised that the price premium for m7a isn't larger than it is, given that you're getting twice as many physical CPU cores. I don't use AWS. For anyone that does... how much does it cost to specify 1 thread per core? Is that a potential optimization that a whole bunch of people are missing?