Benchmarking Amazon's Graviton2 Performance With 64 Neoverse N1 Cores Against Intel Xeon, AMD EPYC

[phoronix]

Phoronix: Benchmarking Amazon's Graviton2 Performance With 64 Neoverse N1 Cores Against Intel Xeon, AMD EPYC

This week Amazon announced the general availability of their EC2 "M6g" instances powered by their second-generation Graviton processors. Amazon is offering a variety of M6g instances with the Graviton2 CPU, including a bare metal instance. In this article are many benchmarks looking at the various Amazon EC2 M6g instances compared to other EC2 Intel Xeon and AMD EPYC instances as well as looking at the M6g Graviton2 metal performance up against various Intel/AMD CPUs in our lab.

http://www.phoronix.com/vr.php?view=29183

[milkylainen]

Indeed impressive results, although it is very interesting that bare metal tests paint a very different picture.

But to be fair, 32 cores and 64 threads are not the same as 64 cores.
Is this Amazon overselling / underselling products to gain foothold?

[tildearrow]

Graviton2 m6c CPUs

Hmmm... I don't see any m6c there :<
I only see m6g...

[schmidtbag]

Shockingly good results, especially for ARM.

[edwaleni]

Shockingly good results, especially for ARM.

At least Amazon makes it publicly available for testing along with Ampere eMag. Can't say that about Marvell ThunderX2.

[torsionbar28]

It would be useful to compare the IPC (instructions per clock) of x86 CPUs vs Graviton in order to determine whether the different instruction encoding and the slightly different operational semantics of instructions is decisive in terms of performance: If Graviton's IPC is overall lower than x86 IPC in case of those benchmark pairs for which the runtimes are roughly the same, then ARM's ISA (instruction set architecture) is more efficient than x86. Or vice versa.

Agreed but that's only one piece of the puzzle. Core topology, pipeline length, core width, branch predictor effectiveness, cache size and architecture, memory bandwidth, and platform I/O, are all significant factors in overall system performance. After all, even within x86_64 designs, we see significant performance deltas when comparing AMD vs intel, even on same process node with same core count and clock speeds. In the end, ISA efficiency being a few percents different one way or the other is probably negligible.

At first glance, these Gravitron2 numbers are impressive. Unfortunately for the ARM fans out there, I don't think this will ever translate into commercial products you can buy. I think Amazon is keeping this as an in-house technology, to give AWS competitive advantage through reduced dependency on AMD and intel. But I guess if it creates a market for ARM servers, this is a net positive.

[bug77]

Man, the first time AMD corners the lucrative server market, Amazon goes ahead and does this...

[Marc Driftmeyer]

What's with these odd RAM sizes on EPYC?

62, 124, 248?

[Marc Driftmeyer]

Agreed but that's only one piece of the puzzle. Core topology, pipeline length, core width, branch predictor effectiveness, cache size and architecture, memory bandwidth, and platform I/O, are all significant factors in overall system performance. After all, even within x86_64 designs, we see significant performance deltas when comparing AMD vs intel, even on same process node with same core count and clock speeds. In the end, ISA efficiency being a few percents different one way or the other is probably negligible.

At first glance, these Gravitron2 numbers are impressive. Unfortunately for the ARM fans out there, I don't think this will ever translate into commercial products you can buy. I think Amazon is keeping this as an in-house technology, to give AWS competitive advantage through reduced dependency on AMD and intel. But I guess if it creates a market for ARM servers, this is a net positive.

And then Zen 3 arrives in a few months making these numbers very unimpressive.

[MrEcho]

Im very impressed with the numbers. And I could see them being used on a lot of different types of workloads. Maybe a lot of internal workloads like DynamoDB, and other internal services.
What I would be worried about is noisy neighbors hammering the ram, and your instance is suffering. Its why you need all those memory channels for VM servers.
It doesnt happen too too often you get a noisy neighbor on AWS, but when a physical host gets maxed out, you know it.