AMD uses two 64-core Epyc server CPUs to smash Cinebench R23 world record

[Shawn Knight]

Recap: For those not familiar, the Epyc 7763 is AMD’s latest top-end server chip. Each one boasts 64 cores with 128 threads in an 8x8 core configuration. With two of them, we’re looking at a whopping 128 cores and 256 threads. Oh, and did I mention they cost nearly $8,000 each?

AMD last month rolled out its third-gen Epyc Milan-based CPUs for servers. The platform has already set dozens of new world records, but what are they capable of when running software that’s a bit more mainstream?

In an attempt to answer this very question, AMD recently brought two of its monster Epyc 7763 chips together to have a go at Cinebench R23, a popular platform for testing a computer’s hardware capabilities.

AMD was able to smash the record without much effort with a score of 113,631. The test was ru-run to include a CPU-Z window in the screenshot and in this test, the score was 113,566. That’s the official one submitted to HWBot.

So, what does this tell us? Well, not much really, except perhaps serving as a bit of a teaser for the sort of performance we can expect from AMD’s next-gen Threadripper lineup.

Found is a TechSpot feature where we share clever, funny or otherwise interesting stuff from around the web.

Permalink to story.

https://www.techspot.com/news/89363-amd-uses-two-64-core-epyc-server-cpus.html

 

[QuantumPhysics]

OK...

How well does it run DCS World ? I don't use benchmarks.

 

[envirovore]

OK...

How well does it run DCS World ? I don't use benchmarks.

Asking someone how it runs a game is using a benchmark.

Also, weren't you the only person here insisting the site use MSFS 2020 for GPU benchmarks?

Why would you even care if you don't use benchmarks?

 

[punctualturn]

 

[punctualturn]

Maybe this one is more appropriate...

 

[VitalyT]

Intel trying to catch AMD...

And it is not for the lack of trying...

 

[Geralt]

OK...

How well does it run DCS World ? I don't use benchmarks.

I always buy the best hardware I can. I am not a fan of Intel or AMD. You should do that too. Instead of having that old-technology-based Core i9, it would be much better for you to have a Ryzen or a Threadripper. AMD is king now. Long live to the king!

 

[deksman2]

OK...

How well does it run DCS World ? I don't use benchmarks.

These are SERVER chips (for data centers), they are not intended for gaming.
If you're looking for gaming benchmarks, you have to look at Ryzen 5700x, 5900x and 5950x.

5950x has the highest core count (16 cores and 32 threads)... and technically surpasses Intel in single core performance (and smashes it in multi-core performance).
However, some games don't do as well on such high core CPU's (the performance drops a bit because the games don't know what to make of so many cores - but not too much).

If you're only gaming, I'd be looking at Ryzen 5600x or 5700x.

If you're doing content creation and gaming, then 5700x, 5900x or 5950x.

Sticking with Ryzen (Zen 3) is better than going with Intel, because AMD is more efficient and generally faster than Intel.
It also gives you better upgrade options, so in case you want to upgrade to Zen 3+ (Warhol), you will probably have that option (as long as the motherboard manufacturer releases a BIOS/UEFI update).

 

[Angga B]

I always buy the best hardware I can. I am not a fan of Intel or AMD. You should do that too. Instead of having that old-technology-based Core i9, it would be much better for you to have a Ryzen or a Threadripper. AMD is king now. Long live to the king!

Until you roam the market to find out there's no Ryzen 9 or TR40 cpu available. Which I did :sad:

Come on AMD and TSMC, help us!

 

[bravesgood]

OK...

How well does it run DCS World ? I don't use benchmarks.

Well...

I don't play DCS World. I run a server..

 

[Ludak021]

So they paired up two cpus to beat one cpu. Am I missing something?

 

[Avro Arrow]

That's a phenomenal score, no doubt. My only question is... What was the previous record and what hardware was used to set it?

 

[Avro Arrow]

So they paired up two cpus to beat one cpu. Am I missing something?

That depends, what hardware was last used to set the record? There's nothing in this article indicating that the previous record was set with only a single CPU. In fact, I'd be shocked if it were considering that dual-socket EPYC and Xeon CPUs were already extant. Where did you get the information that the previous record was set with only a single CPU?

I stand corrected, it was a single TR-3990X. That was unexpected, to say the least.

 

[Tom Yum]

So they paired up two cpus to beat one cpu. Am I missing something?

The article doesn't provide a lot of detail, however the fastest Intel score was from a pair of the latest Ice-Lake based Platinum 8380 processors, and it got a 74630 pts. The previous record was from a heavily overclocked 3990X which got 105170 pts. The 7763's are running stock, so it is impressive in that regard.

 

[Avro Arrow]

The article doesn't provide a lot of detail, however the fastest Intel score was from a pair of the latest Ice-Lake based Platinum 8380 processors, and it got a 74630 pts. The previous record was from a heavily overclocked 3990X which got 105170 pts. The 7763's are running stock, so it is impressive in that regard.

Ah, so it WAS set by a single CPU! It makes sense that it was a 3990X because the best silicon that AMD has actually goes into Threadrippers to increase clock speeds, something that is more or less irrelevant to data centre server chips that rely on maximum throughput and performance-per-watt.

 

[Lionvibez]

Well...

I don't play DCS World. I run a server..

lol QP is totally out of his environment on this topic.

Nobody is playing games on server....

Ah, so it WAS set by a single CPU! It makes sense that it was a 3990X because the best silicon that AMD has actually goes into Threadrippers to increase clock speeds, something that is more or less irrelevant to data centre server chips that rely on maximum throughput and performance-per-watt.

Best silicon goes to Epyc first then Threadripper. Both are higher binned than the consumer Zen 3 chips.

The margins are higher on Epyc also so it takes priority. You are right higher clocked chips go to TR.

 

[Arbie]

Asking someone how it runs a game is using a benchmark.

We are supposed to empathize with QP's simple need to diss AMD at every opportunity. And when there's no opportunity, to just say anything. Here the topic is data center chips running Cinebench for fun - so he complains that it isn't a game. For chrissake... it does get old.

 

[mbk34]

Wouldn't it be nice if the article said what the previous record was? Maybe point out why that benchmark is useful. It might also be interesting to see how a single core compares to a 5600X core, does it scale well using all the cores, all the threads? can you add more than two CPUs? It all just seems a bit like lazy reporting.

 

[Avro Arrow]

Best silicon goes to Epyc first then Threadripper. Both are higher binned than the consumer Zen 3 chips.

The margins are higher on Epyc also so it takes priority. You are right higher clocked chips go to TR.

I've never read that. I've read the exact opposite:
AMD Ryzen Threadripper "Summit Ridge" Dies are Heavily Binned
Now, to be fair, this article is four years old but I've never seen evidence anywhere that AMD has changed their methods. This actually makes sense because higher binning of silicon based on clock speed capability would be completely worthless to EPYCs but not Threadrippers. Binning is all about achieving higher clock speeds and server chips have no need for that. Quite the contrary, clocking a CPU higher than is necessary would result in more power being used and that's a massive negative for a data centre.

If they were to use chips that require more juice, they would want it to be because the CPU had even more cores, not higher clock speeds. If server CPUs were to be binned, it would most likely be based on their ability to run stable while undervolted, not the ones that are capable of higher clocks.

The cost of server CPUs is irrelevant because the market is completely different. You're not buying better silicon with those higher prices, you're buying extremely high levels of support in case anything goes wrong because downtime = lost revenue. That level support is both incredibly extensive and incredibly expensive.

Now, I could be wrong and it could have changed but I haven't read anything to this effect in all of these years. If you could provide a link that proves otherwise, I would be happy to stand corrected and learn something that I didn't know before.