Before I go off and do a whole bunch of testing which could be for nothing, does anyone run a 16 core processor and have completed Michael's testing to get the best thread/task result?

Someone may have done all the hard work and can say that 4*4 works best.
I'm not up for reinventing the wheel :D

The current TDP challenge reignites my liking of PG :D Especially now that I have the settings worked out (Sun, Mercury etc)
____________



Слава Україні!

Thanks Keith.

Lets see if someone comes along. SoB is a pretty specific Thread.
____________



Слава Україні!

Hi Mackerel,

I checked and saw a max FFT of 3200K.
Do I have the "Run FFT's in place" checkbox ticked?

Also, how long would I run the test for?
"Time to run each FFT size"


Thank you

____________



Слава Україні!

Oops. Thank you. And thank you Azmodes.
Can you please explain these results?

http://www.primegrid.com/forum_thread.php?id=8266&nowrap=true#123869

I understand the 1 w/u across 16 cores etc right down to 16 wu's with 1 core each but which is the most productive?
Somewhere in the middle I guess.


____________



Слава Україні!

[Feb 7 23:20] Benchmarking multiple workers to measure the impact of memory bandwidth

[Feb 7 23:20] Timing 16K all-complex FFT, 16 cores, 1 worker. Average times: 0.14 ms. Total throughput: 7306.25 iter/sec.

[Feb 7 23:21] Timing 16K all-complex FFT, 16 cores, 4 workers. Average times: 0.05, 0.06, 0.05, 0.06 ms. Total throughput: 75110.17 iter/sec.

[Feb 7 23:22] Timing 16K all-complex FFT, 16 cores, 8 workers. Average times: 0.06, 0.06, 0.08, 0.07, 0.06, 0.06, 0.06, 0.06 ms. Total throughput: 122487.72 iter/sec.

[Feb 7 23:23] Timing 16K all-complex FFT, 16 cores, 16 workers. Average times: 0.09, 0.09, 0.09, 0.09, 0.09, 0.10, 0.09, 0.10, 0.10, 0.10, 0.09, 0.09, 0.10, 0.10, 0.09, 0.10 ms. Tot

[Feb 7 23:24] Timing 18K all-complex FFT, 16 cores, 1 worker. Average times: 0.15 ms. Total throughput: 6796.79 iter/sec.

[Feb 7 23:25] Timing 18K all-complex FFT, 16 cores, 4 workers. Average times: 0.06, 0.07, 0.07, 0.06 ms. Total throughput: 59685.85 iter/sec.

[Feb 7 23:26] Timing 18K all-complex FFT, 16 cores, 8 workers. Average times: 0.07, 0.09, 0.08, 0.08, 0.08, 0.10, 0.08, 0.08 ms. Total throughput: 97323.04 iter/sec.

[Feb 7 23:27] Timing 18K all-complex FFT, 16 cores, 16 workers. Average times: 0.11, 0.11, 0.11, 0.11, 0.11, 0.11, 0.11, 0.11, 0.12, 0.11, 0.11, 0.11, 0.11, 0.11, 0.11, 0.11 ms. Tot
stopped.

So, the answer is here but I don't know what I'm looking at lol.
Which one is best?

Thanks guys

____________



Слава Україні!

Whoops.

Timings for 16K all-complex FFT length (16 cores, 1 worker): 0.14 ms. Throughput: 7306.25 iter/sec.

Timings for 16K all-complex FFT length (16 cores, 4 workers): 0.05, 0.06, 0.05, 0.06 ms. Throughput: 75110.17 iter/sec.

Timings for 16K all-complex FFT length (16 cores, 8 workers): 0.06, 0.06, 0.08, 0.07, 0.06, 0.06, 0.06, 0.06 ms. Throughput: 122487.72 iter/sec.

Timings for 16K all-complex FFT length (16 cores, 16 workers): 0.09, 0.09, 0.09, 0.09, 0.09, 0.10, 0.09, 0.10, 0.10, 0.10, 0.09, 0.09, 0.10, 0.10, 0.09, 0.10 ms. Throughput: 170700.31 iter/sec.

[Fri Feb 07 23:25:19 2020]
Timings for 18K all-complex FFT length (16 cores, 1 worker): 0.15 ms. Throughput: 6796.79 iter/sec.

Timings for 18K all-complex FFT length (16 cores, 4 workers): 0.06, 0.07, 0.07, 0.06 ms. Throughput: 59685.85 iter/sec.

Timings for 18K all-complex FFT length (16 cores, 8 workers): 0.07, 0.09, 0.08, 0.08, 0.08, 0.10, 0.08, 0.08 ms. Throughput: 97323.04 iter/sec.

Timings for 18K all-complex FFT length (16 cores, 16 workers): 0.11, 0.11, 0.11, 0.11, 0.11, 0.11, 0.11, 0.11, 0.12, 0.11, 0.11, 0.11, 0.11, 0.11, 0.11, 0.11 ms. Throughput: 142896.52 iter/sec.

This is always just a trade off isn't it? 1 core/wu might have a larger throughput number but the result will be slower to complete and therefore less 1st's?


____________



Слава Україні!

Whoops.

Those results aren't helpful for SoB, you forgot to change the FFT size it was testing. 16k and 18k are really tiny. Set it to 3200k for min and max, and see what results that give. You might also add 2 workers into the list of workers it tests, so it would be 1, 2, 4, 8, 16.

What you are looking for from a productivity perspective is higher iter/sec. However, if you want to also be "first" you might sacrifice some throughput for faster units by using a higher number of cores per worker/task.

I've just got the 16 core (32 thread) AMD 1950X and am wondering the same. I've not really done any testing other than I briefly ran 1 x 32 threads, 2 x 16 threads and 2 x 8 threads to see how my temps looked, and 2 x 8 threads ran the most hot (~68C) so I'm taking that configuration to be doing the most computation and am thus using that for now. I've only just built my PC and yet to install decent RAM and do the XMP thing. I suppose that because the CPU has two chiplets I expect either 4 x 4 threads or 2 x 8 threads to be best bet?

I’m thinking the same. I’ll need to do some testing but I think 2 * 8 will be the go. 4*4 will work but I’m still looking for a few 1st and I don’t think I’ll get many with 4*4.
You suffer the same fate as me... thermal throttling at 68 degrees. My system never stays below 67.
____________



Слава Україні!

As I don't have my machine on all the time I'm going to stick with 2x8 so they finish as quick as reasonably possible yet still running hopefully efficiently, and should get a few firsts too :)

I've not overclocked anything, while SOB is running the CPU runs at 4000Mhz (40x100MHz), my board is an X570 Aorus Master.

Indeed I do have a 3950X :) sorry if I caused any confusion. What the best configuration for me was on my mind when I saw the thread pop up and failed to read the title properly.

Thanks for the advice, I will look into it.

I'll be quiet now :)