Der8auer breaks dual-socket record!

[S w a t s o n]

So looks like AMD has now taken the record for both Single and Dual Socket, at least for the time being. Der8auer was able to reach a maximum clock of 4.04GHz on 2x Epyc 7601 32 Core Processors for a score of 10535.

Oh note, it's a novec cooled sever rack and it's going to be shown at Gamescom!

Personally I think this is some next level shit, he mentions it almost beating a quad socket score at some point as well. I can't wait to see what he does with Epyc 2

[Froody129]

We're gonna have to get something more intensive than Cinebench. At this point it's passing with unstable clocks because the CPUs are just so fast that it's just a burst. Consumer level is still great 

[S w a t s o n]

Just now, Froody129 said:

We're gonna have to get something more intensive than Cinebench. At this point it's passing with unstable clocks because the CPUs are just so fast that it's just a burst. Consumer level is still great 

Cinebench will probably just have to release a new render for the benchmark scaled for these high core count CPUs. Part of the problem is finding a benchmark that is threaded and optimized for an arbitrary number of cores at 100% usage

[TechyBen]

Does Blender scale up to 32 cores?

Many render engines do... but then without optimisation, most of these things are pointless. They hit cache limitations and "bottlenecks".

 

Servers can be optimised, or just share resources... and depending on workload, will have different cache requirements or latency requirements anyhow.

[lewdicrous]

An absolute madman!

[Taf the Ghost]

39 minutes ago, Swatson said:

Cinebench will probably just have to release a new render for the benchmark scaled for these high core count CPUs. Part of the problem is finding a benchmark that is threaded and optimized for an arbitrary number of cores at 100% usage

There was 11.5 and then 15. We'll probably get a Cinebench R19 Benchmark in the future, especially as they improve multi-threading performance, since the Industry has finally kicked off the Core Wars.

[GoldenLag]

49 minutes ago, Taf the Ghost said:

There was 11.5 and then 15. We'll probably get a Cinebench R19 Benchmark in the future, especially as they improve multi-threading performance, since the Industry has finally kicked off the Core Wars.

aCOREding to what sources?

 

Its all about threading the needle with all these EPYC news.

 

Its hard to cache everything to keep up. 

[Taf the Ghost]

2 minutes ago, GoldenLag said:

aCOREding to what sources?

 

Its all about threading the needle with all these EPYC news.

 

Its hard to cache everything to keep up. 

[Spotty]

20 minutes ago, GoldenLag said:

aCOREding to what sources?

 

Its all about threading the needle with all these EPYC news.

 

Its hard to cache everything to keep up. 

Those are some Epyc puns.

[GoldenLag]

3 minutes ago, Spotty said:

Those are some Epyc puns.

Why have you Ryzen to the occasion? Why do you have to be Athlon again?

What intel have gathered from your whiskylake?

[cj09beira]

thats crazy, though i would like to see more benches like super pi, gpu pi cpu etc

[LukaH]

if linus won't do anything with epyc i'm glad that der8auer will.

[S w a t s o n]

So to be clear, the world record overall for Cinebench is about 11.5k on 4x Intel Platinum 8180 28-core Xeons. This got 10.5k with 2x Epyc 7601.

 

 

[leadeater]

7 minutes ago, Swatson said:

So to be clear, the world record overall for Cinebench is about 11.5k on 4x Intel Platinum 8180 28-core Xeons. This got 10.5k with 2x Epyc 7601.

Serve the Home also said at these scales and performance Cinebench is basically broken/unreliable as a benchmark, cool to look at but we need something tougher now. It's a bit like taking the 100m race and making it 1m then trying to measure the fastest person in the world, if you're able to walk you'll post a good time heh.

[Doobeedoo]

That's really awesome though. 

[S w a t s o n]

1 minute ago, leadeater said:

Serve the Home also said at these scales and performance Cinebench is basically broken/unreliable as a benchmark, cool to look at but we need something tougher now. It's a bit like taking the 100m race and making it 1m then trying to measure the fastest person in the world, if you're able to walk you'll post a good time heh.

Eh, that would imply that any CPU would break cinebench, it's clear that we need a new one but until then it's a valid comparison, if it's broken for intel it's probably broken for epyc

[leadeater]

Just now, Swatson said:

Eh, that would imply that any CPU would break cinebench, it's clear that we need a new one but until then it's a valid comparison, if it's broken for intel it's probably broken for epyc

Yea it's broken for both, the issue is to do with the run length of the test because clock jitter and rounding has a bigger impact than the actual performance. If you took one CPU out of that quad 8180 it'll post the same score.

[S w a t s o n]

1 minute ago, leadeater said:

Yea it's broken for both, the issue is to do with the run length of the test because clock jitter and rounding has a bigger impact than the actual performance. If you took one CPU out of that quad 8180 it'll post the same score.

Now that's a far stretch, I don't think that would be the case. I haven't seen an tri-socket scores because who runs only 3/4 sockets lmao. I can definitely see it not scaling well but I think having 28 less cores would impact the scores, it is making use of the cores. Previous quad socket record was like 10k back when 24 core was the top dog

[S w a t s o n]

8 minutes ago, leadeater said:

Yea it's broken for both, the issue is to do with the run length of the test because clock jitter and rounding has a bigger impact than the actual performance. If you took one CPU out of that quad 8180 it'll post the same score.

Dual 8180 is like ~6500 so 11.5k for 4x 8180 seems like reasonable scaling to me

[leadeater]

8 minutes ago, Swatson said:

Now that's a far stretch, I don't think that would be the case. I haven't seen an tri-socket scores because who runs only 3/4 sockets lmao. I can definitely see it not scaling well but I think having 28 less cores would impact the scores, it is making use of the cores. Previous quad socket record was like 10k back when 24 core was the top dog

Scaling isn't the issue CB can use all the cores as it is right now, the issue is the run length specifically. It's far to short to get a reliable reading so it's not useful to actually compare which system is better/higher performance, it's completely lost in the error margin.

 

It takes time to schedule threads and allocate work which happens at the start but the test finishes so quickly a large portion of the run time is initialization and not computation.

 

Quote

One area you will note is that the entirety of runs is in the 5-7 second range. That is important as modern CPUs will typically hit an all core turbo mode. Along those lines, we are seeing fairly massive differences, at times greater than 20% peak to valley on some runs. Generally when we see that level of variance, and consistent variance (e.g. not one of 100 runs but every run moving significantly), we know it is time to take a look at a benchmark.

 

As an example, even though our c-ray “hard” setting we developed in 2012 is starting to run into the same 5-7 window, it is still producing repeatable runs with well under 5% benchmark variance. We are still going to be introducing our 8K version soon simply to get longer run times on large machines.

 

5-7 seconds is an extremely short time to run a “benchmark” on a modern CPU. We have run significantly longer workloads on this machine, the types that take days to run and they are extremely consistent. Even tasks like doing large compile jobs in linux are predictable where we have a sub 1% test variance over 100 runs. 20% is enormous in comparison.

 

Quote

We are making the suggestion that Maxon increase the test render scene size. From what we can see, the benchmark is pushing work to all 224 threads. At the same time, with such a short runtime and extremely inconsistent results, the workload needs to run longer to make any initialization negligible. In the professional rendering industry, people do not optimize for 6-7 second renders. It is the multi hour and day (sometimes longer) renders that creative professionals are trying to reduce.

https://www.servethehome.com/cinebench-r15-is-now-a-broken-as-a-benchmark-and-11-5k-surpassed/

 

Any benchmark with a consistent 20% variance is unreliable.

[S w a t s o n]

Just now, leadeater said:

Scaling isn't the issue CB can use all the cores as it is right now, the issue is the run length specifically. It's far to short to get a reliable reading so it's not useful to actually compare which system is better/higher performance, it's completely lost in the error margin.

 

It takes time to schedule threads and allocate work which happens at the start but the test finishes so quickly a large portion of the run time is initialization and not computation.

 

 

https://www.servethehome.com/cinebench-r15-is-now-a-broken-as-a-benchmark-and-11-5k-surpassed/

Sure but my point is that removing a single 8180 from a quad socket system would not yield the same score. Assuming 3.5k per 8180 (which is generous, more like 3.25k) you'd be looking at 10.5k at best with perfect scaling and assuming the benchmark wasn't broken. Clearly being at only 11.5k shows the benchmark is having problem but it's still scaling. It appears to mostly be an issue of variance.

 

14 minutes ago, leadeater said:

Yea it's broken for both, the issue is to do with the run length of the test because clock jitter and rounding has a bigger impact than the actual performance. If you took one CPU out of that quad 8180 it'll post the same score.

 

[leadeater]

8 minutes ago, Swatson said:

Sure but my point is that removing a single 8180 from a quad socket system would not yield the same score. Assuming 3.5k per 8180 (which is generous, more like 3.25k) you'd be looking at 10.5k at best with perfect scaling and assuming the benchmark wasn't broken. Clearly being at only 11.5k shows the benchmark is having problem but it's still scaling. It appears to mostly be an issue of variance.

 

 

Dual 8176 is 6500 ish, Dual 8180 is 7300-7400 ish, but even then it varies at lot. The run time is too short even for these dual socket systems.

 

If the test has 20% error and 1 CPU is 25% of total CPUs in system then taking one out will give basically the same scores, within 5%-10%. And yes I know that is bad math, but the point is the error is too high to get any real meaning out of the test and that is due to the run length.

[S w a t s o n]

3 minutes ago, leadeater said:

Dual 8176 is 6500 ish, Dual 8180 is 7300-7400 ish, but even then it varies at lot. The run time is too short even for these dual socket systems.

 

If the test has 20% error and 1 CPU is 25% of total CPUs in system then taking one out will give basically the same scores, within 5%-10%.

But we're not talking about averages we're talking about max score to max score. A 3 socket 8180 system that scaled perfectly would max out at 10.5k. A quad socket system in the real world has hit 11.5k. That's it, thats proves that removing one 8180 would lower the score.

 

The fact that it has variance down to -20% doesn't matter from a world record perspective

 

Dual stock 8180 is about 6500 from what I can see, 7300 for an OC which linus got when he turned it in the BIOS

[leadeater]

1 minute ago, Swatson said:

But we're not talking about averages we're talking about max score to max score. A 3 socket 8180 system that scaled perfectly would max out at 10.5k. A quad socket system in the real world has hit 11.5k. That's it, thats proves that removing one 8180 would lower the score.

 

The fact that it has variance down to -20% doesn't matter from a world record perspective

I think you missed the point, the quad socket and triple socket systems CAN get the same scores which SHOULD not be possible, that's why its unreliable. If it were longer the score could be higher, we don't know because it's too short to get an accurate measurement.

[S w a t s o n]

Just now, leadeater said:

I think you missed the point, the quad socket and triple socket systems CAN get the same scores which SHOULD not be possible, that's why its unreliable. If it were longer the score could be higher, we don't know because it's too short to get an accurate measurement.

I didn't miss the point, I took issue with your claim. I pointed out in the OP that the DUAL socket epyc almost reached the quad socket score.