AMD EPYC 9965 "Turin Dense" Delivers Better Performance/Power Efficiency vs. AmpereOne 192-Core ARM CPU

[phoronix]

Phoronix: AMD EPYC 9965 "Turin Dense" Delivers Better Performance/Power Efficiency vs. AmpereOne 192-Core ARM CPU

Complementing the AMD EPYC 9575F / 9755 / 9965 performance benchmarks article looking at those Turin processors up against prior AMD EPYC CPUs and the Intel Xeon competition, this article is looking squarely at the 192-core EPYC 9965 "Turin Dense" processor compared to Ampere Computing's AmpereOne A192-32X flagship processor. It's an x86_64 vs. AArch64 battle at the leading 192 core count for performance and CPU power efficiency.

https://www.phoronix.com/review/amd-epyc-9965-ampereone

[Kjell]

Ampere was founded in 2017 and today they're producing ARM processors which aren't far away from catching up to AMD's flagship CPU

Things are slowly taking shape with ARM performance in Linux continuously improving and more distros enhancing support for it

Really can't wait to buy a ARM powered workstation in the future

[ddriver]

What are the "perf per cost" figures based on? The CPU price in isolate? The amortized cpu price in the complete system cost?

I'd say that it is not at all illustrative, and can be quite misleading metric to pretend the CPU is this standalone thing which you buy for the list price and you use it. You put that thing into a mobo, into a chassis, adding memory, storage, compute, periphery, pay people to service and use it.

A vast difference in cpu price can translate into a minor difference in operational cost. It is not a "this cpu is 100% more expensive and 100% faster" - useless since you can't run the cpu on itself. It is a much easier call - "this system is 20% more expensive and 100% faster".

Now, if there's an actual double difference in operational cost, that's a whole different thing.

[Michael]

What are the "perf per cost" figures based on? The CPU price in isolate? The amortized cpu price in the complete system cost?

I'd say that it is not at all illustrative, and can be quite misleading metric to pretend the CPU is this standalone thing which you buy for the list price and you use it. You put that thing into a mobo, into a chassis, adding memory, storage, compute, periphery, pay people to service and use it.

A vast difference in cpu price can translate into a minor difference in operational cost. It is not a "this cpu is 100% more expensive and 100% faster" - useless since you can't run the cpu on itself. It is a much easier call - "this system is 20% more expensive and 100% faster".

As noted in the footnote of the graph, it's the CPU price.

TCO server price is too difficult to equate / subject to much fluctuation depending upon the server OEM/ODM.... Including that Turin being on a reference server without a set price. And then in the case of AmpereOne Supermicro server, they still haven't shared any total server price with me nor is it listed in their eStore yet.

[NateHubbard]

today they're producing ARM processors which aren't far away from catching up to AMD's flagship CPU

Well, currently the only advantage they seem to have is cost, assuming you could actually buy one.
Maybe if they did offer a decent workstation or desktop, it could be interesting.

[JayBee]

Why such an emphasis on idle power? Are these high CPU count server class systems idle often? I would not think so.
It would be helpful to go through the bios settings for each as well as the OS settings such as the governor profiles used, etc. There are often options in the bios and the OS to tune towards performance and or power management. How much work is done by phronix to ensure the operating configuration is as close to each other as possible?

[Gamer1227]

Yep, ARM lacking in servers, what a surprise.

Sticking to 128 bit vectors is not doing them any favors, when Zen5 cores have 512 bit vectors even in the compact variant, huge difference in HPC, Ampere need chiplets to buff the cores.

[ejolson]

TCO server price is too difficult to equate / subject to much fluctuation depending upon the server OEM/ODM....

If one compares list prices of previous generations: A two-socket Epyc server with total 128 cores is about $10,000 more than a two-socket Altra with the same number of cores and similar amounts of network, RAM and storage. The difference in CPU list prices of Zen 5c and AmpereOne seem similar.

For cloud-style microservices AmpereOne may be a good fit. What I would appreciate in these reviews of high-core-count CPUs is a scaling analysis that determines how much a busy VM affects the performance of the other VMs. Presumably the reason AmpereOne has such slow cores is to avoid memory-bandwidth bottlenecks that might impact VM service-level guarantees on multi-tenant systems.

[trifud]

The AmpereOne CPUs aren't optimized for high performance. They are good for running web servers, let's say. Not for LULESH, HPCG, GROMACS, OpenFOAM...

[oleid]

Michael
How can the joules per run be lower while the average watts are higher for Epic?
This sounds to me as if the averaging is not done correctly.
If the average is properly weighted (i.e. the individual watt measurements are properly sampled), the "average watts" times "duration of the test" corresponds to the "joules per run".