A Look At The AMD EPYC Performance On The Amazon EC2 Cloud

[phoronix]

Phoronix: A Look At The AMD EPYC Performance On The Amazon EC2 Cloud

Of the announcements from yesterday's AMD Next Horizon event, one that came as a surprise was the rolling out of current-generation EPYC processors to the Amazon Elastic Compute Cloud (EC2). Available so far are the AMD-powered M5a and R5a instance types to offer Amazon cloud customers more choice as well as being priced 10% lower than comparable instances. Here are some initial benchmarks of the AMD performance in the Amazon cloud.

http://www.phoronix.com/vr.php?view=27061

[s_j_newbury]

Michael in the system info on the first page it says "Full generic retpoline" instead of "Full AMD retpoline", won't this hurt performance?

[Michael]

Michael in the system info on the first page it says "Full generic retpoline" instead of "Full AMD retpoline", won't this hurt performance?

Not sure for how Amazon configures EC2... I am using the default kernel on Ubuntu 18.04 LTS on EC2 and just reporting what it's saying. I assume Amazon/AWS has it in their optimized configuration.

[edwaleni]

Epyc 7571

16c/32t

PS7571BDVIHAF

https://www.amd.com/Documents/AMD-Product-Master.pdf

[Buntolo]

How did you test the EPYC 7571 if it isn't publicly available yet?

[Jaxad0127]

How did you test the EPYC 7571 if it isn't publicly available yet?

As the article states, Amazon Web Services received them before retail availability. I can confirm that EC2 shows them.

[amehaye]

Intel will generally win in media-related benchmarks because AMD's SIMD implementation has about 1/2 the bandwidth IIRC. AMD said that this is going to be fixed in the Zen 2 CPU architecture.

[xception]

Would have loved to see how the cloud compares to standalone EPYC cpus in performance in the same article. I think it would make it easier for people to figure out if it's preferrable for them to buy the hardware or rent the processing power depending on how much each task requires that processing power and for how long.

P.S. I don't actually need that kind of processing power for anything, just thought it would be useful for others.

[Spring]

https://openbenchmarking.org/embed.p...ha=9e619be&p=2 - the annotations are wrong on this graph, clearly those are not "seconds".

And the article also says "With the Parboil OpenMP MRI Gridding benchmark, the M5A instances were leading over EPYC." as if M5A were not Epyc instances. Or what's more probable the labels are in reverse order somehow.

[msroadkill612]

As a newb, some modern twists favouring upstart (reduced historical barriers to entry) Amd seem to be: a/ the cloud & b/ huge improvement tho it is, Zen2 is a simple drop in replacement for existing Epyc platforms (as is Zen3 afaik) (such an advance from intel would require a new platform afaict?).

A prospect for in house hardware, can do a relaxed validation of his app on the cloud & ignore the intel FUD, & existing epyc users have already done the hard yards of validating for epyc on "pilots" based on zen1. There is no need for hesitancy on rollouts of Zen2 epyc.