Zen 3 speculation

[Xyzzy]

We ran a very convoluted benchmark yesterday to determine how well the CPU scales with additional cores. It would take forever to explain what we did so we will just post the "efficiency" results.

I5 = 10600K
R5 = 5600X
Both at default "STK" power levels and with the same memory settings.

   |  I5  |  R5
1C | 100% | 100%
2C |  96% |  97%
3C |  92% |  93%
4C |  87% |  92%
5C |  78% |  87%
6C |  72% |  83%

[M344587487]

AMD's SMT implementation is reportedly better than intel's but it may just be that AMD handles MT better in general. If your test benefits from SMT it would be interesting to see how the 12C results compare to the 6C results.

[mackerel]

Originally Posted by Xyzzy

We ran a very convoluted benchmark yesterday to determine how well the CPU scales with additional cores. It would take forever to explain what we did so we will just post the "efficiency" results.

Even without full details, can I take it the CPUs were allowed to boost as far as they like? I wonder how it would have looked at fixed clock. What FFT size was tested? I suspect that may be the contributing factor here.

Originally Posted by M344587487

AMD's SMT implementation is reportedly better than intel's but it may just be that AMD handles MT better in general. If your test benefits from SMT it would be interesting to see how the 12C results compare to the 6C results.

Define "better"? I only did in depth testing on Zen and Zen+, never finished it on Zen 2 and no plans to buy a Zen 3. For Zen and Zen+ they had difficulty extracting core utilisation when running a single thread through it, even though (outside of FP work) they had more general resource than Intel. Thus they crept ahead once you put two threads through them. So they had more improvement, but starting from a relatively lower position.

I have to give up on if HT/SMT helps with P95 type work. Old testing says no, newer testing says... maybe. But it is rather narrow. I did see scenarios where using more threads can recover performance from bad configuration elsewhere, but in that scenario it doesn't actually get you any more performance, just less loss.

Testing by another user elsewhere suggests Zen 3 might have about 10% higher IPC in P95 like uses over Zen 2, where not limited by ram etc.

[M344587487]

Better as in the workloads that do benefit from SMT tend to have a bigger improvement on AMD than intel, reportedly as in this is based on something I've consumed within the last year and can't remember the source so pinch of salt. You're right that it could be that AMD's single thread performance was lacking and SMT picked up the slack. There's plenty of isolated on vs off data but nothing I can find offhand comparing on vs off with Zen2/3 vs intel all at once, well there is some but it seems exclusively game testing which almost always is better with SMT off on all architectures so it's not really relevant.

[BigDukeSix]

Just ran a short benchmark of actual FFTs on a 5600X@stock on a ASUS TUF GAMING B550M PLUS board with 32 GB RAM Dualchannel @ 1600 MHz according to CPU-Z. I did the benchmark mainly to see how one worker performes in comparising to 2 workers. I think, one worker is the way to go.
But I don´t know if it is a problem, Prime95 doesn´t know ZEN 3 yet and so has no CPU optimization yet.

Prime95 64-bit version 30.3, RdtscTiming=1
Timings for 5376K FFT length (6 cores, 1 worker):  3.28 ms.  Throughput: 304.92 iter/sec.
Timings for 5376K FFT length (6 cores, 2 workers):  7.98,  8.13 ms.  Throughput: 248.38 iter/sec.
Timings for 5760K FFT length (6 cores, 1 worker):  3.91 ms.  Throughput: 255.64 iter/sec.
Timings for 5760K FFT length (6 cores, 2 workers):  9.61,  9.64 ms.  Throughput: 207.77 iter/sec.
Timings for 6144K FFT length (6 cores, 1 worker):  4.22 ms.  Throughput: 236.96 iter/sec.
Timings for 6144K FFT length (6 cores, 2 workers): 10.34, 10.44 ms.  Throughput: 192.47 iter/sec.
Timings for 6400K FFT length (6 cores, 1 worker):  4.39 ms.  Throughput: 228.03 iter/sec.
Timings for 6400K FFT length (6 cores, 2 workers): 10.85, 10.86 ms.  Throughput: 184.25 iter/sec.
Timings for 6720K FFT length (6 cores, 1 worker):  4.54 ms.  Throughput: 220.24 iter/sec.
Timings for 6720K FFT length (6 cores, 2 workers): 11.27, 11.28 ms.  Throughput: 177.39 iter/sec.
Timings for 6912K FFT length (6 cores, 1 worker):  4.96 ms.  Throughput: 201.48 iter/sec.
Timings for 6912K FFT length (6 cores, 2 workers): 12.40, 12.51 ms.  Throughput: 160.64 iter/sec.
Timings for 7168K FFT length (6 cores, 1 worker):  4.85 ms.  Throughput: 206.27 iter/sec.
Timings for 7168K FFT length (6 cores, 2 workers): 11.86, 12.30 ms.  Throughput: 165.65 iter/sec.
Timings for 7680K FFT length (6 cores, 1 worker):  5.26 ms.  Throughput: 190.03 iter/sec.
Timings for 7680K FFT length (6 cores, 2 workers): 12.58, 12.66 ms.  Throughput: 158.50 iter/sec.
Timings for 8064K FFT length (6 cores, 1 worker):  5.67 ms.  Throughput: 176.38 iter/sec.
Timings for 8064K FFT length (6 cores, 2 workers): 13.54, 13.53 ms.  Throughput: 147.80 iter/sec.
Timings for 8192K FFT length (6 cores, 1 worker):  5.78 ms.  Throughput: 173.02 iter/sec.
Timings for 8192K FFT length (6 cores, 2 workers): 14.53, 14.26 ms.  Throughput: 138.92 iter/sec.

[Xyzzy]

Originally Posted by BigDukeSix

But I don´t know if it is a problem, Prime95 doesn´t know ZEN 3 yet and so has no CPU optimization yet.

We emailed George a special debug file so he can add the information.

[BigDukeSix]

I like it

[zurek]

Searching this forum, I found this to be the only thread dealing with Prime95 and AMD Rzyen 5950x. I was trying to run Prime95 on this processor (with Asus Dark Hero, set at 44.5 on both CCX overclock).

It passes all the torture tests, except the small FFT. When I run this test, computer crashed immediately after the test is started (some 2 seconds after the test button hit). After restart, bios reports overheated CPU. But given that the test barely started, it seems too early for the CPU to get overheated. And since other tests are OK, I assume, there might be a bug in Prime95, also due to the fact (I learned on this forum), that Prime95 still does not know Zen3!


Also, other stress testing software (AIDA and others) are OK on my machine.


When I run Prime95 and do some torture tests, I also noticed in HWinfo, that Prime95 does not respect the values set in bios for overclocking, and runs the CPU at much higher speeds! I do not think this is correct. My CPU is set to 4450MHz max in BIOS overclocking, while Prime95 reports the CPU being 4725.05 MHz (CPU setting and Information) which is obviously a wrong figure.


Hence, I assume that the computer crash is not because of the instability of my machine, but because of a bug in Prime95. We need a new version of Prime95, capable of dealing correctly with Zen 3 architecture. Am I right?

[mackerel]

Zurek, based on what you have written it sounds like your computer is unstable. Small FFT is one of if not the most intense load you can put on a CPU core. Large FFTs are limited by connectivity bandwidth. Aida64 if you select FPU/cache only might come close, but if you mix other selections it is a much less intense test.

Not having specific support for Zen 3 doesn't really matter. If the clock is reported wrong that is probably just a mis-reading but it doesn't actually change the running clock. Verify using another tool like hwinfo.

To check: turn off the overclock and see if the system is stable. If you really want to overclock, try PBO since it is a power limit based OC. Adjusting clock/voltage manually is not ideal since Zen 2 as workloads can vary a lot in their demand.

[retina]

Zurek, P95 doesn't, and can't, change the CPU frequency. That is what the OS/BIOS does.

And two seconds for overheating is definitely possible. The power can ramp up very quickly and the cooling can't always keep up. You need to look into your cooling setup, and also look into your OS/BIOS settings to keep the clock frequencies in check.

Your hardware has a problem, fix it. Wondering if P95 has a bug is not a solution to fix hardware. If P95 has a bug then it would crash, or produce an incorrect result, or not run, or something similar. It can't cause the BIOS to detect overheating unless the CPU really is overheating, or the BIOS has a bug, or both.

[LaurV]

If you see higher clocks when P95 runs, what I suspect is that the "Turbo" features are kicking in.

When your computer does nothing, the clock is reduced to basic, to save energy.

When some program or process needs work to be done, the clock is raised to some higher limit, called "Turbo mode".

Turbo mode is not supposed to run continuously, or at all, without a very good cooling.

As the other posters said, P95 can not change your clocks.

And heating the CPU "to red hot" in seconds is quite normal if the heat transfer is not proper. Think about you put a hundred watt on few square centimeters. Those centimeters will smoke instantly. That is what coolers are for. Thermal paste, big aluminum or copper fins, large fans, water pumps... we won't invest a lot of money into them if they won't be needed. Their job is to take the heat away. Otherwise the magic smoke will go out from the components, and all components that run on smoke will cease to work.

Can you reduce the clocks to some lower limit, and see if the FFT test passes? If yes, no bug. If not, it may be related to some P95 bug. There was a case in the past, but George fixed it already.