Prime95 30.7

[Micke]

*edit*
I just don't use this bugged old version anymore, which is listed as actual version on https://www.mersenne.org/download/#download.

[tshinozk]

I run mprime on linux (Mint) in the same machine, in turn.

Timings for 2048K FFT length (1 core, 1 worker): 0.47 ms. Throughput: 2119.90 iter/sec.

mprime still uses all cores.
This result corresponds to the case of 12900k in post #203.

[tshinozk]

My BIOS (X299) can not disable AVX-512, while it seems RocketLake can.
I found that "noxsave" kernel parameter disables AVX.
https://stackoverflow.com/questions/...linux-computer

I run linux with "noxsave" for the boot option of GRUB.
As a result, all AVX are disabled, and Throughput shows very slow for SSE.
mprime still uses all cores, while the old version is not so.

Intel(R) Core(TM) i9-7980XE CPU @ 2.60GHz
CPU features: Prefetchw, SSE, SSE2, SSE4

Prime95 64-bit version 30.7
Timings for 2048K FFT length (1 core, 1 worker): 1.14 ms. Throughput: 880.96 iter/sec.
Timings for 2048K FFT length (2 cores, 1 worker): 1.14 ms. Throughput: 879.03 iter/sec.

Prime95 64-bit version 30.6
Timings for 2048K FFT length (1 core, 1 worker): 16.28 ms. Throughput: 61.41 iter/sec.
Timings for 2048K FFT length (2 cores, 1 worker): 8.29 ms. Throughput: 120.60 iter/sec.

[S485122]

Originally Posted by tshinozk

My BIOS (X299) can not disable AVX-512, while it seems RocketLake can.
I found that "noxsave" kernel parameter disables AVX.
https://stackoverflow.com/questions/...linux-computer

I run linux with "noxsave" for the boot option of GRUB.
As a result, all AVX are disabled, and Throughput shows very slow for SSE.
...

It might be easier to tell Prime95 whether you want to use AVX512 features : stop (and exit ?) mprime, edit local.txt.

From undoc.txt :

The program supports many different code paths for PRP/LL testing depending on
the CPU type.  It also has a few different factoring code paths.  You can
force the program to choose a specific code path by setting the proper
combination of these settings in local.txt:
	CpuSupportsRDTSC=0 or 1
	CpuSupportsCMOV=0 or 1
	CpuSupportsPrefetch=0 or 1
	CpuSupportsSSE=0 or 1
	CpuSupportsSSE2=0 or 1
	CpuSupports3DNow=0 or 1
	CpuSupportsAVX=0 or 1
	CpuSupportsFMA3=0 or 1
	CpuSupportsFMA4=0 or 1
	CpuSupportsAVX2=0 or 1
	CpuSupportsAVX512F=0 or 1
This shouldn't be necessary though as the program uses the CPUID instruction
to see if the CPU supports these features.

[kruoli]

It should not be necessary to disable AVX-512 to get a working-as-expected configuration. But I concur with Jacob to try if disabling AVX-512 the way he described leads to the same behaviour.

@tshinozk: Have you had a look on v30.8 already? While is not ready-to-release yet, it is worth having a look at it to check whether it has this bug fixed. If not, you can post to the 30.8 thread that the bug is also present there and I guess it is more likely to get attention there.

[tshinozk]

v30.8b14 has the issue, starting from v30.7b2 .

I try running prime95.exe with CpuSupportsXXX in local.txt .
I can disable AVXs, but the issue is not resolved.

Can nobody reproduce the issue where prime95 cannot change the number of cores in benchmark?

[James Heinrich]

Originally Posted by tshinozk

v30.8b14 has the issue, starting from v30.7b2

As kruoli said, if the problem exists in the current development version you may want to post in the v30.8 thread since that version is currently being worked on (and v30.7 is not).

[kriesel]

Yes.
Mprime/prime95 normally runs at lower priority to other user applications and system tasks. It yields the CPU for those higher priority tasks as much as they can use. With significant system load, benchmarking on available cycles on half the cores or less may look very similar to benchmarking on all of them.
The following is a quick benchmark on my laptop I'm typing on now, which has lots of tabs in Firefox open, and numerous (dozens) remote desktop sessions going as client (display). Firefox alone was using around 2 cores' throughput out of the 4 real on this Windows 10 system.

[May 6 11:13:06] Worker starting
[May 6 11:13:06] Your timings will be written to the results.bench.txt file.
[May 6 11:13:06] Compare your results to other computers at http://www.mersenne.org/report_benchmarks
[May 6 11:13:06] Benchmarking multiple workers to measure the impact of memory bandwidth
[May 6 11:13:07] Timing 2048K FFT, 1 core, 1 worker.  Average times:  5.59 ms.  Total throughput: 179.00 iter/sec.
[May 6 11:13:22] Timing 2048K FFT, 1 core hyperthreaded, 1 worker.  Average times:  6.72 ms.  Total throughput: 148.83 iter/sec.
[May 6 11:13:38] Timing 2048K FFT, 2 cores, 1 worker.  Average times:  5.35 ms.  Total throughput: 187.04 iter/sec.
[May 6 11:13:53] Timing 2048K FFT, 2 cores hyperthreaded, 1 worker.  Average times:  6.63 ms.  Total throughput: 150.94 iter/sec.
[May 6 11:14:09] Timing 2048K FFT, 3 cores, 1 worker.  Average times:  5.58 ms.  Total throughput: 179.11 iter/sec.
[May 6 11:14:25] Timing 2048K FFT, 3 cores hyperthreaded, 1 worker.  Average times:  5.90 ms.  Total throughput: 169.37 iter/sec.
[May 6 11:14:41] Timing 2048K FFT, 4 cores, 1 worker.  Average times:  5.46 ms.  Total throughput: 183.05 iter/sec.
[May 6 11:14:56] Timing 2048K FFT, 4 cores hyperthreaded, 1 worker.  Average times:  5.86 ms.  Total throughput: 170.73 iter/sec.
[May 6 11:15:12] 
[May 6 11:15:12] Throughput benchmark complete.
[May 6 11:15:12] Throughput benchmark complete.
[May 6 11:15:12] Worker stopped.

Those timings are suspiciously similar even given the other system loads. See first two attachments.

So, switching to its hardware twin "martinette", no Firefox running, only 1 VNC remote desktop server running, and prime95 V30.8b14, CPU utilization ~5-10% with prime95 paused, benchmarking 1-4 cores with & without hyperthreading, 1 worker, Windows 11 TaskManager shows all physical cores running saturated regardless of indicated core count during prime95 benchmarking; logical cores' cpu loading % are somewhat affected by hyperthreading or not. (attachments 3 - 5)


Another oddity is that it would not run 8192 fft benchmarking. In v30.7b9 I could specify 2048-2048 as fft lengths to benchmark, 1-4 cores, 1 worker only.

V30.8b14 exited benchmarking instantly, to stopped state, if I specified 8192-8192. As does v30.7b9, for 4096 only or 8192 only. There might be more. 6144 behaved ok.

[kruoli]

Originally Posted by kriesel

See first two attachments.

The first attachment is especially worrysome because main memory and both the SSD are fully utilized. What are "other system loads"? Do they justify this loads? If not, it looks like extensive swapping.

[kriesel]

Originally Posted by kruoli

What are "other system loads"?

Mostly Firefox-gone-wild. I've dumped a lot of tabs which gave very temporary relief. It gets in strange states sometimes if left running a long time, including when it has downloaded a FF update but not yet applied it & restarted FF. I tend to leave it running as long as possible with multiple Google Colab sessions, and many other tabs open. Looks like time for a clean start again RSN.
Checking just now, Help, About Firefox includes "Restart to update Firefox".
And after temporarily taming Firefox, prime95 v30.7b9 demonstrates the same all-physical-cores-busy-when-benchmarking-1 issue as v30.8b14.

[Prime95]

Originally Posted by tshinozk

v30.8b14 has the issue, starting from v30.7b2 .

Try v30.8b15