Intel Xeon Platinum 8490H "Sapphire Rapids" Performance Benchmarks

[Paradigm Shifter]

These servers are 40 grand each, and each processor has 60C/120T and 16 accelerators; that's 120C/240T and 32 accelerators per server.

I think that's a rather conservative price estimate. Perhaps with no storage or very little RAM a dual-60C box would be 40,000. With a reasonable quantity of RAM-per-core, I would image it would be double that. I recently tripled the price of the Intel server I specced up by half-filling its maximum memory capacity. In the end we went with Epyc because it was cheaper, no slower and would arrive faster (and most of what we do is GPU based, so the CPUs are just there to keep the GPUs fed).

If you have a program which can use these new AMX extensions, Sapphire Rapids is a good buy. But in that scenario, it's the only buy, so there we are.

[sophisticles]

So what you're saying is that anyone who wants a general purpose CPU should just buy AMD?

I don't think there is anyone with even a modicum of intellectual honesty that would argue against the fact that at the moment AMD offers the more compelling options if one is interested in running highly threaded applications on general purpose processors.

The point is that these specific processors are not aimed at that demographic, they are aimed at those considering buying a Z15/16 or upgrading from a Z14/15 to a Z16.

These servers are 40 grand each, and each processor has 60C/120T and 16 accelerators; that's 120C/240T and 32 accelerators per server.

Ten of these servers in a rack would cost 400 grand and feature 1200C/2400T and 320 accelerators, which compares very favorably with what is found in a 500 grand Z15.

No one is buying these things to compile code or run blender or run CAD.

That's what makes this review, and the comments regarding AMD superior overall showing, so laughable.

[Keats]

As I said, it's obvious that Michael does not understand the purpose of these processors, but he is clearly not alone as evidenced by the comments from "crudeboil:".

The accelerators found on these chips have specific uses, in compression and decompressing, encryption and decryption, databases (analytics):

https://www.intel.com/content/www/us...analytics.html

https://www.intel.com/content/www/us.../overview.html

https://www.intel.com/content/www/us...-overview.html

https://www.intel.com/content/www/us...l#gs.mc7xob

The Xeon Platinum 8490H are Intel's answer to the processors found in IBM's Z15/16 servers, those processors are basically a collection of accelerators bound together with a general purpose cpu; the sole purpose of the cpu is to boot up the OS and initialize the peripherals, once that's done everything is handled by accelerators, encryption, decryption, math, I/O, compression, decompression, all of it, performed by accelerators, programmed using COBOL.

These Xeons are made for running custom software, much of it coded in Python, running libraries that can leverage the on-board accelerators.

The fact that Michael tested this server in the way he did, and then defending his test, tells me he doesn't understand the purpose of this system and frankly should never have been supplied with a test unit.

So what you're saying is that anyone who wants a general purpose CPU should just buy AMD?

[AdrianBc]

Some of Michael's reviews are great, some are so-so and some prove that he doesn't understand what he is reviewing.

This is one of those reviews.

He is given a high end, purpose built server, with two $17,000 Xeon Platinum 8490H, each with 16 dedicated accelerators, and he runs benchmarks like code compilation.

This is the equivalent of a car reviewer being given a Ford 350 Lariat Super Cab Long Box with a goose neck and doing 0-100-0, skid pad and slalom tests.

No one one is buying a Xeon Platinum 8490H to run any of the software bench marked in this review.

They are buying them to run specialized software that can benefit from the on-board accelerators.

Thus this whole review, any any others like it by other reviews, is basically a waste of time.

I think that it is you who does not understand the purpose of this review.

No individual or small business will waste money by buying 8490H CPUs, but many might buy some of the cheaper Sapphire Rapids models, e.g. 24-core SKUs.

Most Sapphire Rapids SKUs have most of the accelerators disabled, unless the owner chooses to pay for Intel on Demand, which makes sense only for big companies, like cloud vendors. So the Sapphire Rapids accelerators will be of no benefit for individuals or small businesses.

The employees of big companies do not care much about the benchmarks published here, because their companies can afford to buy or rent some samples of Sapphire Rapids systems on which to test exactly the applications in which they are interested.

On the other hand, individuals or small businesses cannot afford to spend thousands of $ for a Sapphire Rapids server, only to discover that it was not worth buying, so they are very interested in published benchmarks like these, especially in code compilation and computational benchmarks like those included in this review.

I am much more interested to see the same benchmarks from this review, but between Xeon 5412U ($1113) and Ryzen 7950X, but the results for 8490H and Genoa are still useful as a proxy enabling a very rough estimation of how those would compare, even if the estimation is hard, due to the great differences between the memory interfaces of Xeon and Ryzen.

[Mahboi]

Good article. But I think Michael you were being very charitable with the comments.

In Michael's defense, at this point Intel deserves our charity. Being rough with these Xeons may be akin to child beating.

[nicalandia]

...

Michael, why do you think there is a Huge performance difference in OpenVINO between the 1P and 2S 8490H?

image.png

image.png

image.png

[nicalandia]

I wouldn't pay too much attention to that page, yet, due to its limited scope. It's only showing data where there is sufficient collection for each benchmark/configuration and other factors that refines it to only showing sound generalized results. As more benchmarks of the 8490H and other SPR results appear in the days/weeks/months ahead it will be much more relevant.

I am using the released data that you just published

2S 9374F(64C/128T) is beating a single 8490H even with the OpenVIMO AMX numbers.

image.png

AMD Genoa 2S 9374F vs 1P Intel Xeon SPR 8490H

[sophisticles]

My theory is that nobody will buy these 8490H for QAT.

Then you do not understand what QAT is nor the history of what made Intel a powerhouse.

Back in the day, there used to be what was known as Big Iron, powerful servers running UNIX, that consumed a lot of power, cost a ton of money and had used UNIX OSes that had extremely expensive per user licensing costs.

Wintel was born when Intel adopted the "fast enough" marketing campaign, the sold processors that were half as fast as the cpu's in these Big Iron servers, used a tenth of the power and cost a fraction of what the huge servers cost. MS for its part, released Windows version with relatively cost effective licensing terms, this Wintel partnership is what is credited with bringing Big Iron down.

Intel is clearly going down the same road again, these Xeons despite costing 17 grand each, are way cheaper than the minimum 500K to 3000K for the IBM Z15.

So the sales pitch is obvious, they are marketing a processor with 16 on-board accelerators, with the implication that a rack of dual xeon servers is still cheaper than the Z15 main frame.

The Z16's are another matter, supposedly they are so fast that they are actual quantum safe.

[willmore]

Michael (?) Typo on page 5: "2P match0-up."

[crudeboil]

My theory is that nobody will buy these 8490H for QAT.