AmpereOne A192-32X Benchmarks: 192 Core ARM Server Performance & Power Efficiency

[phoronix]

Phoronix: AmpereOne A192-32X Benchmarks: 192 Core ARM Server Performance & Power Efficiency

Last week an AmpereOne server finally arrived at Phoronix! Ampere Computing sent over a reviewer server of the AmpereOne A192-32X flagship AArch64 server processor with 192 custom cores and using a Supermicro ARS-211M-NR R13SPD platform. I have been carrying out a number of benchmarks for this much-anticipated AArch64 cloud native processor and have initial performance and power efficiency metrics to share today to see how it compares to prior Ampere Altra Max as well as the Intel Xeon and AMD EPYC server competition.

https://www.phoronix.com/review/ampereone-a192-32x

[Markopolo]

To start with I’d like to thank you for this awesome review, and linking the pts results - that is always appreciated!

I do have one recurring criticism though. I think it would be better if you made and followed a standardized format for your benchmark reviews. The problem is that you often change the order of benchmarks, or even worse, which benchmarks you include in the review itself and which get omitted. This approach is made even worse when it comes to power and efficiency numbers. As a reader it is rather jarring, and difficult not to feel like you’re unnecessarily trying to spin a narrative even if I know that’s not your intent (all the data seems to be there in the linked pts results)

[Michael]

To start with I’d like to thank you for this awesome review, and linking the pts results - that is always appreciated!

I do have one recurring criticism though. I think it would be better if you made and followed a standardized format for your benchmark reviews. The problem is that you often change the order of benchmarks, or even worse, which benchmarks you include in the review itself and which get omitted. This approach is made even worse when it comes to power and efficiency numbers. As a reader it is rather jarring, and difficult not to feel like you’re unnecessarily trying to spin a narrative even if I know that’s not your intent (all the data seems to be there in the linked pts results)

It pretty much is standardized. The exception(s) though are like with AArch64 where a number of benchmarks aren't supported on AArch64 or otherwise tuned for it or similar obstacles. And then for some benchmarks like when reviewing Genoa-X / Xeon Max will obviously focus more on looking at workloads where the cache can be interesting, etc etc. But then with the OB link all of the results are there for those interested. The same goes for the power consumption and all the perf-per-Watt results being on the OB link. I don't show them for each one all the time as otherwise it gets rather repetitive and would make the articles exceedingly long.

Edit: the main exception I can think of off the top of my head is when running cloud benchmarks and for the instances when I am running cloud benchmarks out of pocket, I run far fewer benchmarks in order to reduce my costs. Or cases where I am in a time crunch.

[S.Pam]

Thanks for the review! Interesting to see more competition in the server market. I am a bit disappointed that the new Arm isn't that efficient on the wattage side of things, but perhaps that will improve as platforms develop further?

What I would like to see added to the reviews here on Phoronix is the current price for the CPUs included. While the CPU price isn't the same as the overall platform cost, it still makes it easier to follow whether a CPU is more or less valuable compared to the others in each benchmark.

[Michael]

Thanks for the review! Interesting to see more competition in the server market. I am a bit disappointed that the new Arm isn't that efficient on the wattage side of things, but perhaps that's coming.

What I would like to see added to the reviews here on Phoronix is the current price for the CPUs included. While the CPU price isn't the same as the overall platform cost, it still makes it easier to follow whether a CPU is more or less valuable compared to the others in each benchmark.

Unfortunately price comparisons to be accurate on the server CPU side is rather difficult. Unless just going off the list price but even then isn't too relevant. Particularly with large organizations/hyperscalers typically getting deep discounts on CPUs, CPU pricing can vary widely from retail to ordering as part of server from OEM/ODM, lack of any public/retail availability right now for AmpereOne and Sierra Forest, etc.

But for those wanting to compare prices on their own or have a particular CPU supplier, clicking the OB link at end of article at top of page users can fill in their desired pricing and get all the perf-per-dollar graphs too.

[guspitts]

Interesting benchmarks.

When it comes to power consumption, I also wish the total amount of Joules would be reported, in addition to the average Watts. In other words a way to answer the question: do the faster processors solve the problem using less total energy because they solve it faster (even though it is using higher wattage while processing)? Or how much more energy does one use for the benefit of faster computation?

[Michael]

Interesting benchmarks.

When it comes to power consumption, I also wish the total amount of Joules would be reported, in addition to the average Watts. In other words a way to answer the question: do the faster processors solve the problem using less total energy because they solve it faster (even though it is using higher wattage while processing)? Or how much more energy does one use for the benefit of faster computation?

As a footnote on each power graph for individual benchmarks is the joules used for that given benchmark.

[guspitts]

As a footnote on each power graph for individual benchmarks is the joules used for that given benchmark.

Ah, yes, I see it now. I must be too used to skip over the fine prints... Thanks.

[S.Pam]

Unfortunately price comparisons to be accurate on the server CPU side is rather difficult. Unless just going off the list price but even then isn't too relevant. Particularly with large organizations/hyperscalers typically getting deep discounts on CPUs, CPU pricing can vary widely from retail to ordering as part of server from OEM/ODM, lack of any public/retail availability right now for AmpereOne and Sierra Forest, etc.

But for those wanting to compare prices on their own or have a particular CPU supplier, clicking the OB link at end of article at top of page users can fill in their desired pricing and get all the perf-per-dollar graphs too.

Perhaps add c/t@hz on each cpu in the chart? It isnt super easy to remember what each model is.

[Yalok]

Thanks for mentioning idle power consumption. In real world scenarios the server is not usually under high load 100% of time.
So it might not be that good of a deal if it consumes x10 power when idle over competition from Intel or AMD.

I've seen people noticing high idle power consumption with desktop zen4 CPUs over desktop Intel Gen 13. But I don't remember you mentioning this in your desktop zen4 tests.