NVIDIA Vera CPU Benchmarks: Olympus Cores Delivering The Best Performance Ever Seen On ARM Review

Posted by Artoriuz@reddit | hardware | View on Reddit | 40 comments

Across the variety of different workloads that were permitted for testing on this initial round of NVIDIA Vera benchmarking, Vera exceeded my expectations in never seeing an ARM64 processor compete so well against the x86_64 competition. On a geo mean basis, the NVIDIA Vera delivered 10% better performance than the AMD EPYC 9575F 5.0GHz high frequency processor. For gen-on-gen compared to Grace, Vera was coming in at 1.63x the performance geo mean. Over a single Intel Xeon 6980P as Intel's current flagship Granite Rapids processor, NVIDIA Vera delivered 1.55x the performance.

[-]

Noble00_@reddit

Also as a disclaimer:

NVIDIA also requested only specific workloads relevant to the intended workloads/domains that Vera is catering to in the data center be tested. So this first round of Vera benchmarking isn't too comprehensive across the spectrum of possible workloads but limited to the benchmarks that were permitted based on what they feel were most relevant -- plus the fact I was only spending one day at NVIDIA's offices. For these initial NVIDIA Vera benchmarks they preferred the scope of benchmarks be limited to target use-cases they feel most relevant for their modern data center customers. This isn't a sponsored article but I obliged to their requests in order to run these initial Vera CPU benchmarks. Hopefully in future rounds of Vera testing over the months ahead there will be a more widespread set of benchmarks for those curious about the Olympus CPU core performance at large. Similarly, in being able to report on the power efficiency and performance-per-Watt once their power management code is tuned.

Nevertheless, Nvidia is making a splash into the CPU DC market. It blows past the rest when it comes to memory bandwidth. It'll be interesting to see power metrics when that is available. AMD's Venice and Intel's Seirra Forrest is right around the corner (though, Diamand Rapids seems to be coming late) and it'll be interesting to see the performances of these DC products head to head.

[-]

Exist50@reddit

which Vera competes with 2P sockets

The 2P numbers in the article are barely better than 1P, so I don't think that comparison makes sense.

[-]

mckirkus@reddit

Yes, best to wait for more benches but the AMD EPYC 9575F is a monster of a CPU. This think might be a CFD chart topper. The AMD MI300 also stuffs a bunch of Epyc cores onto the package with HBM, so it seems the architectures are all moving in that direction.

[-]

mennydrives@reddit

Granted, caveats applied, with NVIDIA limiting the scope of the initial benchmarks

Yeah, I'm just gonna take this with a coffee-mug-sized grain of salt for now.

I love Phoronix but that kinda shoulda been front and center.

[-]

TheBraveOne86@reddit

I knew someone would complain. But it’s totally fair. This is a niche product for niche uses. It’s not going to be for gaming, etc.

They said “test our products in scenarios it might be used”.

It’s like a sports car manufacturer asking the auto reviewer not to test towing capacity. Well duh. Sure that Porsche might be able to tow a boat. But it wasn’t really designed to, so if you say it sucks at towing a boat are you really reviewing it appropriately?

[-]

Whirblewind@reddit

And this is called apologism.

[-]

ColdStoryBro@reddit

Phoronix doesnt test server cpus on games. They have a big suite of productivity and enterprise apps. No llama.cpp, openvino apache...

[-]

Noble00_@reddit

No llama.cpp, openvino apache...

But they do? https://www.phoronix.com/review/amd-epyc-9745-9755/3

[-]

Exist50@reddit

That said, their suite is kind of bad in general.

[-]

Noble00_@reddit

It's on the front page.

[-]

mennydrives@reddit

In the very last paragraph of the first page. It should honestly be in the headline, a sub-headline, or the opening paragraph, not at the bottom of the page.

NVIDIA also requested only specific workloads relevant to the intended workloads/domains that Vera is catering to in the data center be tested.

[-]

Noble00_@reddit

Honestly I feel it's like a litmus test on this sub when these types of articles are shared. People just go straight to the final page read the chart and be done with it.

Sure it could be the 1st paragraph, but if you actually read it, it's simply that of introductory writing then once the background and specs are written down, Michael writes whatever is necessary to say (mind you, it's not like the information is lost in the middle of page, it's just at the closing).

[-]

Daftpunk67@reddit

I mean it’s fairly obvious that the scope of testing was going to be limited anyways when they mention they were invited to nvidias hq. I’m sure intel and AMD do the same things as well, and there’s probably more as well, but as a first look at the product I don’t really see an issue.

[-]

Noble00_@reddit

I agree. It just irks me with the way people consume content these days. The article already tells you it's a first look, performance captured at vendor specific environment and yet the reader acknowledges but still argues how it's 'unfaithful'.

[-]

LAwLzaWU1A@reddit

How dare you suggest that people should actually read an article?!

I think it is funny that someone a few replies above says this info should be in the headline. Maybe they think it should be included in the headline because that is the only thing they read?

[-]

dagmx@reddit

Did I miss where this article actually compares it to other arm cores? All the actual comparisons are to x86…

[-]

trololololo2137@reddit

server arm chips are pretty bad in general. i think they would win easily against neoverse

[-]

WinterCharm@reddit

AWS runs on their custom Gravirtron chips and they're very cost effective and quite good.

[-]

trololololo2137@reddit

per core perf sucks ass compared to arm phone chips. arm is not using their best tech for servers

[-]

bb999@reddit

Per core performance doesn’t matter for server, cost per performance is what matters

[-]

nittanyofthings@reddit

Are you sure the benchmark used in that comp is relevant? Mobile benchmarks don't test sustainable loads, just burst.

[-]

DerpSenpai@reddit

Neoverse cores all mostly on E cores or older P cores

[-]

Artoriuz@reddit (OP)

Yes, Grace has Neoverse V2 cores (roughly equivalent to Cortex-X3) . That's very old tech by now.

[-]

thehighshibe@reddit

cortex x3 old?? 😭 _clutches cortex a7 closely_

[-]

Rodot@reddit

Didn't they compare to neoverse (GH superchips)?

[-]

marcost2@reddit

Well this was a heavily sponsored piece with nVidia even restricting the benchmarks they could run, so not surprised

[-]

Noble00_@reddit

Early look + Nvidia is breaking into the data center CPU market, so that's why it's compared to Intel/AMD

[-]

Extreme-Arm4609@reddit

This is awesome I cannot wait for this to be in laptops cuz nvidia's obviously going to have custom cores and laptops it's probably not going to be exactly this but it's probably going to be close to this.

Now when are we going to see it I'm guessing like 2028-2027 something like that.

N1 series is going to be standard arm designs intograted by Mediatek, N2 probably going to be like the tegra series custom made by Nvidia but with Nvidia designed cores.

Literally everyone right now is going to make custom core Samsung Apple Qualcomm have/will soon. That's the hot new thing to do.

So this is kind of interesting for that and also the PPW completely unconfirmed because we were just looking at CPUs running at way past the efficiency peak they are basically running at vmax but I'm not sure. So it's kind of worthless to look at that and Nvidia forbid them from looking at actual performance per watt and frequency scaling because it's not finished yet in software.

Super sad because that's the interesting part.

[-]

Extreme-Arm4609@reddit

Dann y'all fucking hate everything I say lol.

I just uh, idgaf anymore

[-]

Late_Scarcity3455@reddit

Is this the same arch the N1/N1X will be based on?

[-]

Extreme-Arm4609@reddit

Nope N1 is standard last gen arm cortex cores.

This is probably what N2 will be based on, because remember this pairs with Rubin in the data center, that was paired with Blackwell and Blackwell in the data center used normal arm cores just like N1 does.

[-]

Slasher1738@reddit

no, I believe N1/N1x is based on Grace using the NeoVerse cores

[-]

tnoy@reddit

N1/N1X is apparently just based on the GB10, so it's going to be 10 Cortex-X925 and 10 Cortex-A725 cores.

[-]

Artoriuz@reddit (OP)

No. Those have standard ARM cores.

[-]

EloquentPinguin@reddit

Nope, N1X will use the same CPU clusters as we know:

5+5 Cortex X925

5+5 Cortex A725

[-]