News Dual AMD EPYC 7742 Crushes Quad Intel Xeon Platinum 8180M In Geekbench 4

[Admin]

ServeTheHome benchmarks two AMD EPYC 7742 and compares them to four Intel Xeon Platinum 8180M.

Dual AMD EPYC 7742 Crushes Quad Intel Xeon Platinum 8180M In Geekbench 4 : Read more

 

[MasterMadBones]

By now we know that Geekbench results for Zen 2 CPUs are totally unrepresentative of real world performance. It just benefits too much from the new cache hierarchy. I don't understand why Tom's is still posting these results with every new CPU.

 

[jeremyj_83]

By now we know that Geekbench results for Zen 2 CPUs are totally unrepresentative of real world performance. It just benefits too much from the new cache hierarchy. I don't understand why Tom's is still posting these results with every new CPU.

What are you talking about? While benchmarks like Geekbench aren't great, the Zen 2 architecture is almost universally revered for its performance.

 

[MasterMadBones]

What are you talking about? While benchmarks like Geekbench aren't great, the Zen 2 architecture is almost universally revered for its performance.

It is, but outliers like these can give the wrong impression. Indeed, I don't doubt that even in more reliable benchmarks, the dual 7742s will pull out a significant lead and there are plenty of reviews to back that up.

 

[jeremyj_83]

It is, but outliers like these can give the wrong impression. Indeed, I don't doubt that even in more reliable benchmarks, the dual 7742s will pull out a significant lead and there are plenty of reviews to back that up.

There are a lot of reviews that backup that the 7742 is the best, and overall the Epyc 7002 series as well, server CPU you can buy. I am interested in more of the reviews of the 7552 to add some 2U4N hosts to the data center I administer.

 

[thGe17]

By now we know that Geekbench results for Zen 2 CPUs are totally unrepresentative of real world performance ...

Couldn't agree more. I've checked Geekbench results many times for several CPUs and im my opinion, they only indicate a tendency. And also there are some sub-tests, which strongly correlate with L3 cache size.
The next question: why have they chosen to show the 8180M? It's an older Skylake model and additionally it's the more expensive M-version.
The current Cascade Lake-SP has minor optimizations and is produced in an optimized process (higher clock, same TDP) and therefore a better comparison. And again, the plain 8280 is already sufficient, because in a 4-way-system, you can already put 4 TiB RAM into the system (so no need to cherry pick its L-version for about 17000 $ unless you have a special need for it).
Looking at current SPEC2017 results 128 Rome-cores are approximately on a par with 112 Cascade Lake-cores. In INTrate, the results are almost the same, in FPrate, the Intel-Systems are (almost insignificantly) faster.
Beyond SPEC, if special applications make heavy use of AVX-512, Rome ist still no match for Intel, but these are no general purpose (server) apps (e. g. NAMD and GROMACS).

Aside from that it's also obvious, that Rome currently has the better price/performance ratio, but in the end, price is not the only factor that matters, especially in a data center. With Rome and its lower pricing, decision making gets a little bit easier, but its not that simple that as of today the answer will always be "Epyc/Rome".
AMD also knows that. L. Su stated in the last earnings call something like "SKU prices are a secondary concern in data centers" and F. Norrod expects that AMD gains 10 % market share (IDC) at the earliest by end of 2019 or until mid 2020. AMD has done their part and now its up to partners and OEMs to do their part, building boards, servers, the whole ecosystem.

 

[jeremyj_83]

Couldn't agree more. I've checked Geekbench results many times for several CPUs and im my opinion, they only indicate a tendency. And also there are some sub-tests, which strongly correlate with L3 cache size.
The next question: why have they chosen to show the 8180M? It's an older Skylake model and additionally it's the more expensive M-version.
The current Cascade Lake-SP has minor optimizations and is produced in an optimized process (higher clock, same TDP) and therefore a better comparison. And again, the plain 8280 is already sufficient, because in a 4-way-system, you can already put 4 TiB RAM into the system (so no need to cherry pick its L-version for about 17000 $ unless you have a special need for it).
Looking at current SPEC2017 results 128 Rome-cores are approximately on a par with 112 Cascade Lake-cores. In INTrate, the results are almost the same, in FPrate, the Intel-Systems are (almost insignificantly) faster.
Beyond SPEC, if special applications make heavy use of AVX-512, Rome ist still no match for Intel, but these are no general purpose (server) apps (e. g. NAMD and GROMACS).

Aside from that it's also obvious, that Rome currently has the better price/performance ratio, but in the end, price is not the only factor that matters, especially in a data center. With Rome and its lower pricing, decision making gets a little bit easier, but its not that simple that as of today the answer will always be "Epyc/Rome".
AMD also knows that. L. Su stated in the last earnings call something like "SKU prices are a secondary concern in data centers" and F. Norrod expects that AMD gains 10 % market share (IDC) at the earliest by end of 2019 or until mid 2020. AMD has done their part and now its up to partners and OEMs to do their part, building boards, servers, the whole ecosystem.

The reason for the 8180M comparison is that result was already in the Geekbench database, Servethehome only ran the benchmark of the dual 7742s. The original review they did of the 7742 shows that outside of AVX512 code, the dual 7742 is typically 80-100% faster than the top end Platinum. With AVX512 the dual 7742s still edged out dual Platinums, but not by much. You should read their review it is quite extensive and informative. https://www.servethehome.com/amd-epyc-7002-series-rome-delivers-a-knockout/

 

[thGe17]

The reason for the 8180M comparison is that result was already ...

Yes, that's possible, but in this case if they wanted to, they could have omitted the price column. But that's not solely a tomshw problem, more a general media an advertising phenomenon to create "more important, more impressive news".
Of course Epyc is about 80 - 100 % faster and I would be surprised otherwise. It has approx. the same core-throughput, therefore with more than twice the number of cores it must be around twice as fast ;-) Plain and simple math.
With servethehome you already have the proper link and as you can see, a 8280 (28 C) and also a 8260 (24 C *) is faster (with heavy loaded AVX-512) than a 7742 (64 C). Only the proper dual socket setup enables Rome to close the gap and it finally "wins" this benchmark with less than +1,0 % (utilizing 128 cores vs. only 56 cores). And almost the same picture in AnandTech's review, Here, a Dual 8280 "outperforms" a Dual 7742 in NAMD with +1,9 % (again 56 vs 128 cores). This is interesting, because it often puzzled me how effective AVX-512 could be in real (but very specialized) applications, considering the fact, that Intel has to reduce its core frequency significantly for AVX-512 code.

Don't get me wrong, Rome is a marvelous piece of hardware, and everybody should be grateful for AMDs endeavor (even pure Intel-fans). Competition is good for the market.
But as I mentioned before, 42 is not the answer to all questions in life and also Epyc/Rome is not the answer to all data center related questions ;-)

*) Btw, the Platinum 8260 is far less expensive than the 7742.

**) To be fair, with these AVX comparisons in mind there is still a chance that Epyc/Rome might become faster in the future with Rome-specific code optimizations for AVX2 (but it is also obvious that it cannot close the gap to AVX-512).

 

[Alex/AT]

By now we know that Geekbench results for Zen 2 CPUs are totally unrepresentative of real world performance. It just benefits too much from the new cache hierarchy.

The thing is, cache size and performance are extremely crucial to real world tasks, especially enterprise, high level of simultaneous access servers and virtualization tasks. Especially given such high core chips are definitely designed for two of the latter.