AMD EPYC 7351P Single Socket CPU Linux Benchmarks and Review

[Patrick Kennedy]

Our AMD EPYC 7351P linux benchmarks and review includes a comparison to other AMD EPYC CPUs under $1000 and the entire Intel Xeon Silver line

The post AMD EPYC 7351P Single Socket CPU Linux Benchmarks and Review appeared first on ServeTheHome.

Continue reading...

 

[Edu]

Thanks for a great article.
Unfortunately, the power consumption figures are a bit confusing. It says 340W peak power, then 192W for something else (not specified). And then it's not clear which benchmark is running when the 340W peak power is drawn and why do you not measure 100% load?

 

[Patrick]

Thanks for a great article.
Unfortunately, the power consumption figures are a bit confusing. It says 340W peak power, then 192W for something else (not specified). And then it's not clear which benchmark is running when the 340W peak power is drawn and why do you not measure 100% load?

70% load is being shown (192W)
340W peak = 100% AVX2 workload

Remember "100% load" is a bit misleading. There is a large difference these days between 100% integer and AVX2/ AVX-512 load. Also, with how much the cache and RAM are being used when you look at power from a system perspective. Virtually all of our benchmarks are running across each core.

There are a lot of people doing "100% load" testing on parts of the CPU but are missing the memory controllers, and Infinity Fabric links for example.

Treat our maximum numbers as the configuration presented should never exceed that number.

 

[Evan]

Now that looks like solid value for money !! Power consumption aside if you need a powerful since socket system then looks like a winning cpu

 

[Patrick]

Now that looks like solid value for money !! Power consumption aside if you need a powerful since socket system then looks like a winning cpu

Power consumption is not that bad. There are both perf/ watt and machine/U / watt that need to be taken into account.

Xeon Silver does well power consumption/ U.

 

[bitrot]

Sweet Christmas, what a beast of a CPU! Shame about the power consumption trade-off which is rather significant for me personally, but the 7351P pretty much destroys everything Intel has to offer in that price range, as I had expected going by the 4114 benchmarks published recently on here.

I can only imagine how the 7401P benchmarks will look like. Thing is, it's not just that AMD offers a ton of perfomance for a relatively low price, these EPYC P SKUs make lower end dual-socket servers more or less obsolete. If AMD manages to improve / mature the ecosystem and especially improve availability (!), they have a winner for sure.

Great article as always.

 

[Edu]

7

70% load is being shown (192W)
340W peak = 100% AVX2 workload

Remember "100% load" is a bit misleading. There is a large difference these days between 100% integer and AVX2/ AVX-512 load. Also, with how much the cache and RAM are being used when you look at power from a system perspective. Virtually all of our benchmarks are running across each core.

There are a lot of people doing "100% load" testing on parts of the CPU but are missing the memory controllers, and Infinity Fabric links for example.

Treat our maximum numbers as the configuration presented should never exceed that number.

Thanks for the answer.
But which specific benchmark is it drawing the 340W? NAMD, GROMACS, Chess?
Power measurement alone isn't very useful, unless you also measure the performance- then you can measure the efficiency of the processor.

 

[dynamis31]

Impressive article, with an astounding conclusion. I'm amazed to see the level of independance, professionalism and impartialism of this site.
Reviews from STH are being directed towards to my IT staff. These guys are so impregnated with Intel : so many dual socket with only one CPU... time to be rationnal and put saved money in good stuff (storage, network).
Anyway you have the guts to do that !!! Congrats.

 

[tigweld0101]

Power measurement alone isn't very useful, unless you also measure the performance- then you can measure the efficiency of the processor.

Uh....... Unless you're trying to gauge number of nodes per rack. ROFL

Damn son. Nice article. That's a great morning read.

 

[MiniKnight]

Great job @Patrick and the STH team.

 

[Edu]

Uh....... Unless you're trying to gauge number of nodes per rack. ROFL

Damn son. Nice article. That's a great morning read.

Right, but why can't you have both?

 

[Biren78]

Right, but why can't you have both?

From reading STH for years, they've got more data if you want to pay for it. @Patrick

 

[Patrick]

@Biren78 We are publishing less than 2% of our data points publicly these days. Publishing over 1000 performance data points with corresponding power consumption is not feasible.

@Edu - it is an AVX2/ 512 workload that is generating those numbers. Due to NDA we cannot discuss what the software is behind it. We run containers during our test scripts for each CPU for various software vendors. What I can say is that the workload/ power consumption is similar to what we can get up to on the longer GROMACS runs.

Peak values are important because those are generally the ones you worry about either from the perspective of tripping PDU breakers (requiring hands to reset usually) or running into "penalty" style power pricing.

We have idle because let us face it, there are lots of servers running idle.

We have a normal load workload to show what you are going to see on a well-loaded system. That gives us idle, normal heavy load, and worst case. If you have a specific workload you want a number for, we offer DemoEval as a route to test your own workloads on these systems.

 

[Evan]

Even testing your own workload you have to take into account your environmentals that you will host in.... temperature, altitude, humidity, airflow rates, leaving gaps between servers or not, how the room is configured etc, you would be surprised just how much power cooled takes in a server.

For home lab use I am only interested in idle, for general interest or work related topics as @Patrick mentions normal load and peak are important also, but nothing substitutes in house testing.

 

[Edu]

@Biren78 We are publishing less than 2% of our data points publicly these days. Publishing over 1000 performance data points with corresponding power consumption is not feasible.

@Edu - it is an AVX2/ 512 workload that is generating those numbers. Due to NDA we cannot discuss what the software is behind it. We run containers during our test scripts for each CPU for various software vendors. What I can say is that the workload/ power consumption is similar to what we can get up to on the longer GROMACS runs.

Peak values are important because those are generally the ones you worry about either from the perspective of tripping PDU breakers (requiring hands to reset usually) or running into "penalty" style power pricing.

We have idle because let us face it, there are lots of servers running idle.

We have a normal load workload to show what you are going to see on a well-loaded system. That gives us idle, normal heavy load, and worst case. If you have a specific workload you want a number for, we offer DemoEval as a route to test your own workloads on these systems.

That's fair enough. I'm quite surprised that the Epyc draws so much power. It's about 10% faster than the dual 4110, but drawing 70% more power.

 

[Evan]

Intel has for a good number of years been focused on efficiency and the latest 14++ process node or whatever it is in now especially well optimised.

Interested to see the usage on the gold 6130,6140,6150 as a better point of reference.

 

[zir_blazer]

Didn't AMD supported at some point configurable TDP with Opterons?
Details on AMD Bulldozer: Opterons to Feature Configurable TDP

Does EPYC supports anything similar to that? If you capped an EPYC to Xeon Silver levels, how would the performance per watt and price performance of a handicapped EPYC be?

 

[Evan]

Didn't AMD supported at some point configurable TDP with Opterons?
Details on AMD Bulldozer: Opterons to Feature Configurable TDP

Does EPYC supports anything similar to that? If you capped an EPYC to Xeon Silver levels, how would the performance per watt and price performance of a handicapped EPYC be?

Just a feeling is that as with all the current benchmarks the answer is ‘it depends’, you will in general see I think that AMD and Intel IPC this generation is amazingly close !