News Intel's 56-Core Sapphire Rapids Tested: Faster Than 64-Core EPYC

[Admin]

Intel's Xeon Platinum 8480+ 'Sapphire Rapids' beats regular AMD's EPYC 7763, but fails to beat EPYC 7773X.

Intel's 56-Core Sapphire Rapids Tested: Faster Than 64-Core EPYC : Read more

 

[Hooda Thunkett]

I'm wondering how their multi-core results can be so much higher with 56 cores than Epyc at 64 cores if their single core results are so much lower? Can these cores run more than 2 threads each or something?

 

[jeremyj_83]

I'm wondering how their multi-core results can be so much higher with 56 cores than Epyc at 64 cores if their single core results are so much lower? Can these cores run more than 2 threads each or something?

Considering how much faster the 7763X is compared to the 7763 my guess is memory bandwidth. The SPR chip has 8 channel DDR5-4800. That gives it 307GB/sec bandwidth or 5.5GB/sec/core compared to 205GB/sec or 3.2GB/sec/core on the Epyc. With the inclusion of the 3D V-cache the 7763X is able to get around some of the bandwidth constraints seen in the 7763. Similar to how the Infinity Cache in the Radeon 6000 series helps reduce traffic on the VRAM bus so it can use a 256bit bus instead of a 384bit like nVidia.

 

[Makaveli]

useless benchmark however since nobody runs geekbench on a server.

 

[jeremyj_83]

useless benchmark however since nobody runs geekbench on a server.

That is absolutely correct. Geekbench by itself is a garbage metric but places still use it.

 

[TerryLaze]

I'm wondering how their multi-core results can be so much higher with 56 cores than Epyc at 64 cores if their single core results are so much lower? Can these cores run more than 2 threads each or something?

Looks like the xeon was locked at 2Ghz in both single and multi, while the epyc was running 3.5 in single but 2.5 in multi.
Maybe the xeon is still an ES that doesn't have turbos implemented.

If you divide the multi score of the xeon by 56 you get very close to the single core which would confirm that both where run at the same speed, while for the epyc it's way way off.

 

[JamesJones44]

I'm wondering how their multi-core results can be so much higher with 56 cores than Epyc at 64 cores if their single core results are so much lower? Can these cores run more than 2 threads each or something?

Since this is still a limited production test sample the boost parameters might not be correctly optimized yet. It could be turbo boost isn't kicking in for single core at all, but is functional for multicore actions. Until we know more about the specs of the CPU, there are a lot of possible reasons why this could happen.

 

[-Fran-]

This server SoC won't be fighting Milan (or Milan-X for that matter), but Turin and Turin-X (plus the other Zen4 variants). From what we're seeing here (crappy test; agreed, but a test none the less), it's going to be a bloodbath. Intel needed this CPU/SoC out in the wild yesterday.

Regards.

 

[jeremyj_83]

This server SoC won't be fighting Milan (or Milan-X for that matter), but Turin and Turin-X (plus the other Zen4 variants). From what we're seeing here (crappy test; agreed, but a test none the less), it's going to be a bloodbath. Intel needed this CPU/SoC out in the wild yesterday.

Regards.

Considering how ICL only got Intel to Rome they needed SPR released when Milan was released. SPR will still be at a RAM density disadvantage compared to Genoa since AMD is 12 channel vs Intel 8 channel.

 

[Mr_Bond007]

Considering how much faster the 7763X is compared to the 7763 my guess is memory bandwidth. The SPR chip has 8 channel DDR5-4800. That gives it 307GB/sec bandwidth or 5.5GB/sec/core compared to 205GB/sec or 3.2GB/sec/core on the Epyc. With the inclusion of the 3D V-cache the 7763X is able to get around some of the bandwidth constraints seen in the 7763. Similar to how the Infinity Cache in the Radeon 6000 series helps reduce traffic on the VRAM bus so it can use a 256bit bus instead of a 384bit like nVidia.

More likely those results are a load of crap for the Intel or are for a dual socket system. The multicore performance is higher per core than for the single if only 56 cores in the system.

 

[jeremyj_83]

More likely those results are a load of crap for the Intel or are for a dual socket system. The multicore performance is higher per core than for the single if only 56 cores in the system.

You do realize that MT performance can be higher than the equivalent of just 56c due to these having SMT. That means there are 112 threads running and overall SMT can improve performance by about 20%.

 

[Mr_Bond007]

You do realize that MT performance can be higher than the equivalent of just 56c due to these having SMT. That means there are 112 threads running and overall SMT can improve performance by about 20%.

Yes, but that would mean they are scalling the cores linearly, unlikely, but maybe. Also could be that the epyc performance in the table was cherry picked (click bait) for poor multicore performance, for example after a quick look: https://browser.geekbench.com/v5/cpu/15720840
Multithreaded result: 45100, multiple (hundreds) scores found in the 45000 region, only 1 in the 25000 for integer.

 

[TerryLaze]

Yes, but that would mean they are scalling the cores linearly, unlikely, but maybe. Also could be that the epyc performance in the table was cherry picked (click bait) for poor multicore performance, for example after a quick look: https://browser.geekbench.com/v5/cpu/15720840
Multithreaded result: 45100, multiple (hundreds) scores found in the 45000 region, only 1 in the 25000 for integer.

But you can see that the single score is very close so obviously another interpretation could be that the multicore score could have been done in overclock and not with default settings.
And most people that post scores online do it to show off their overclock not to show base performance.

 

[Mr_Bond007]

But you can see that the single score is very close so obviously another interpretation could be that the multicore score could have been done in overclock and not with default settings.
And most people that post scores online do it to show off their overclock not to show base performance.

Every other result for the 7763 is between 41000 and 45000. Thr 7773x in the chart in the article is over 50000. This was purely click bait.

If I get time I will check the performance of one of the dell r7515 that I have at work, with a 7702p.
Found some 7713P that are not in use, will be slower, but should give the performance region.

 

[C_D]

7763 and 5950 results surely are suspicious: the first is too low, the other is too high, comparing to each other. And later, 7773X gives twice multithread performance comparing to 7763, because of 3x cache? There is something wrong with the test, or results are rigged.