Prime95 30.7 - Page 19

storm5510 · 2022-04-30, 23:27

Quote:

Originally Posted by tshinozk

On p95v307b9.win64, I can not chage a number of cores when I run the benchmark.
If I set to 1 core, UI shows 1 core but Prime95.exe uses all cores.

p95v308b13.win64 also has same issue.

You're specifying a range. I don't think it works this way. Just put 18.

kriesel · 2022-05-01, 00:07

Quote:

Originally Posted by James Heinrich

Your screenshot shows 1-18 for "Number of CPU cores to benchmark".

Not a problem. See the note at the bottom of the screen capture's benchmark pane. Prime95 explicitly supports ranges, lists, or lists of ranges, benchmarking successively through 1, 2, 3, ... 18 cores on a single worker by entering 1-18 for example.
Check for other applications using lots of cycles. Firefox can be very CPU and memory intensive.
That can really distort both prime95 benchmark results and what Task Manager CPU monitoring show.

Best benchmarking results will be obtained when all other processes practical are idle or absent.

James Heinrich · 2022-05-01, 00:33

Quote:

Originally Posted by kriesel

Not a problem.

It's only a problem because tshinozk claimed to have entered 1 but got benchmarks (at some point) using all cores.

tshinozk · 2022-05-01, 14:17

I run some old versions.
From the taskmanger or the results of benchmarks, I can distinguish the issue.

p95v306b4.win64 OK (single core)
Timings for 2048K FFT length (1 core, 1 worker): 5.01 ms. Throughput: 199.74 iter/sec.

p95v307b1.win64 OK (single core), but fail to complete
p95v307b2.win64 NG (all cores)
p95v307b3.win64 NG (all cores)
p95v307b4.win64 fail to run, immediately stop
p95v307b5.win64 NG (all cores)
p95v307b7.win64 NG (all cores)
p95v307b8.win64 NG (all cores)
p95v307b9.win64 NG (all cores)

p95v308b13.win64 NG (all cores)
Timings for 2048K FFT length (1 core, 1 worker): 0.57 ms. Throughput: 1741.37 iter/sec.

tshinozk · 2022-05-02, 02:22

It seems that AlderLake has the issue.
The result of 1 core shows too fast, even if AlderLake is running over 5GHz.

12900k:
https://mersenneforum.org/showpost.p...8&postcount=64
Timings for 2048K FFT length (8 cores, 1 worker): 0.62 ms. Throughput: 1602.22 iter/sec.

https://mersenneforum.org/showpost.p...4&postcount=65
FFTlen=2048K all-complex, Type=3, Arch=8, Pass1=128, Pass2=16384, clm=4 (1 core, 1 worker): 0.62 ms. Throughput: 1624.29 iter/sec.

12700K:
https://mersenneforum.org/showpost.p...7&postcount=69
Timings for 2048K FFT length (1 core, 1 worker): 4.57 ms. Throughput: 218.94 iter/sec.
Timings for 2048K FFT length (8 cores, 1 worker): 0.64 ms. Throughput: 1564.59 iter/sec.
It appears that this is normal.

Zhangrc · 2022-05-02, 10:28

Quote:

Originally Posted by tshinozk

It seems that AlderLake has the issue.
The result of 1 core shows too fast, even if AlderLake is running over 5GHz.

12900k:
https://mersenneforum.org/showpost.p...8&postcount=64
Timings for 2048K FFT length (8 cores, 1 worker): 0.62 ms. Throughput: 1602.22 iter/sec.

https://mersenneforum.org/showpost.p...4&postcount=65
FFTlen=2048K all-complex, Type=3, Arch=8, Pass1=128, Pass2=16384, clm=4 (1 core, 1 worker): 0.62 ms. Throughput: 1624.29 iter/sec.

It appears that this is normal.

Maybe all-complex FFT with AVX-512 instruction set is faster.
FFT uses complex numbers, if we compute a complex number directly instead of computing the real and imaginary part respectively, we could get an over 2x speedup.

tshinozk · 2022-05-03, 02:00

"Benchmark all-complex FFTs" option is not much faster as normal in my machine with AVX512.

Timings for 2048K FFT length (1 core, 1 worker): 0.65 ms. Throughput: 1535.94 iter/sec.

Timings for 2048K all-complex FFT length (1 core, 1 worker): 0.65 ms. Throughput: 1546.01 iter/sec.

And both have the issue.(running using all cores)

James Heinrich · 2022-05-03, 02:08

Quote:

Originally Posted by tshinozk

And both have the issue.(running using all cores)

Have you fixed the issue where you're entering 1-18 for "CPU cores to benchmark"? Change that to 1 if you only want to test 1 core...

tshinozk · 2022-05-03, 02:32

No.
Prime95.exe uses all cores, even if I enter 1 in "Number of CPU cores to benchmark" textbox.

Timings for 2048K FFT length (1 core, 1 worker): 0.61 ms. Throughput: 1627.02 iter/sec.
Timings for 2100K FFT length (1 core, 1 worker): 0.75 ms. Throughput: 1338.92 iter/sec.
Timings for 2160K FFT length (1 core, 1 worker): 0.80 ms. Throughput: 1243.79 iter/sec.

Throughput for 1 core is expected to around 100-200 iter/sec for such FFT length

tshinozk · 2022-05-03, 05:08

"FFT timings benchmark" does not have the issue.
I can see the multi-core scaling.

Timing FFTs using 1 core:
Best time for 2048K FFT length: 4.987 ms., avg: 5.001 ms.
Timing FFTs using 2 cores:
Best time for 2048K FFT length: 2.606 ms., avg: 2.891 ms.
Timing FFTs using 3 cores:
Best time for 2048K FFT length: 1.768 ms., avg: 2.066 ms.
Timing FFTs using 4 cores:
Best time for 2048K FFT length: 1.328 ms., avg: 1.853 ms.

tshinozk · 2022-05-04, 06:38

I try reduceing the active cores in BIOS setup.
Even if I activate only 4 cores (disabling 14 cores) and Hyperthread is off , Prime95.exe still uses all cores.

Intel(R) Core(TM) i9-7980XE CPU @ 2.60GHz
CPU speed: 3286.47 MHz, 4 cores
Timings for 2048K FFT length (1 core, 1 worker): 1.42 ms. Throughput: 703.46 iter/sec.

2022-05-01, 14:17	#202
tshinozk Nov 2012 Japan 100011₂ Posts	I run some old versions. From the taskmanger or the results of benchmarks, I can distinguish the issue. p95v306b4.win64 OK (single core) Timings for 2048K FFT length (1 core, 1 worker): 5.01 ms. Throughput: 199.74 iter/sec. p95v307b1.win64 OK (single core), but fail to complete p95v307b2.win64 NG (all cores) p95v307b3.win64 NG (all cores) p95v307b4.win64 fail to run, immediately stop p95v307b5.win64 NG (all cores) p95v307b7.win64 NG (all cores) p95v307b8.win64 NG (all cores) p95v307b9.win64 NG (all cores) p95v308b13.win64 NG (all cores) Timings for 2048K FFT length (1 core, 1 worker): 0.57 ms. Throughput: 1741.37 iter/sec.

2022-05-02, 02:22	#203
tshinozk Nov 2012 Japan 5·7 Posts	It seems that AlderLake has the issue. The result of 1 core shows too fast, even if AlderLake is running over 5GHz. 12900k: https://mersenneforum.org/showpost.p...8&postcount=64 Timings for 2048K FFT length (8 cores, 1 worker): 0.62 ms. Throughput: 1602.22 iter/sec. https://mersenneforum.org/showpost.p...4&postcount=65 FFTlen=2048K all-complex, Type=3, Arch=8, Pass1=128, Pass2=16384, clm=4 (1 core, 1 worker): 0.62 ms. Throughput: 1624.29 iter/sec. 12700K: https://mersenneforum.org/showpost.p...7&postcount=69 Timings for 2048K FFT length (1 core, 1 worker): 4.57 ms. Throughput: 218.94 iter/sec. Timings for 2048K FFT length (8 cores, 1 worker): 0.64 ms. Throughput: 1564.59 iter/sec. It appears that this is normal.

2022-05-03, 02:00	#205
tshinozk Nov 2012 Japan 5×7 Posts	"Benchmark all-complex FFTs" option is not much faster as normal in my machine with AVX512. Timings for 2048K FFT length (1 core, 1 worker): 0.65 ms. Throughput: 1535.94 iter/sec. Timings for 2048K all-complex FFT length (1 core, 1 worker): 0.65 ms. Throughput: 1546.01 iter/sec. And both have the issue.(running using all cores)

2022-05-03, 02:32	#207
tshinozk Nov 2012 Japan 35₁₀ Posts	No. Prime95.exe uses all cores, even if I enter 1 in "Number of CPU cores to benchmark" textbox. Timings for 2048K FFT length (1 core, 1 worker): 0.61 ms. Throughput: 1627.02 iter/sec. Timings for 2100K FFT length (1 core, 1 worker): 0.75 ms. Throughput: 1338.92 iter/sec. Timings for 2160K FFT length (1 core, 1 worker): 0.80 ms. Throughput: 1243.79 iter/sec. Throughput for 1 core is expected to around 100-200 iter/sec for such FFT length