GFLOPS off by factor of 2 on AMD Ryzen 7 5800X 8-Core Processor about peakperf HOT 3 CLOSED

I think I've misunderstood what the printed GLOP value is meant to reflect this whole time. It's not the expected GFLOP/s for my CPU, but rather the number of GLOPs used for each trial. Closing the issue with this comment.

from peakperf.

Let me start by calculating your theoretical peak performance (please correct me if I'm wrong):

PP (SP) = 8 (cores) * 4.0 (Approx. Freq in GHz) * 2 (has FMA) * 8 (AVX2) * 2 (2 x AVX2 Units) = 1024 GFLOP/s

However, you seem to get 2048 GFLOP/s in some cases with 16 threads while you get around 1024 with 8 threads. There are only two possible explanations:

1. My maths are wrong.
2. There is a bug in peakperf.

Can you please paste the complete output of peakperf with no arguments? (which will run using 16 threads).

from peakperf.

Here:

------------------------------------------------------
    peakperf (https://github.com/Dr-Noob/peakperf)
------------------------------------------------------
        CPU: AMD Ryzen 7 5800X 8-Core Processor
  Microarch: Zen 3
  Benchmark: Zen 3 (AVX2)
 Iterations: 1.00e+09
      GFLOP: 2048.00
    Threads: 16

   Nº  Time(s)  GFLOP/s
    1  1.69877  1205.58 *
    2  1.72237  1189.06 *
    3  1.76517  1160.23
    4  1.71196  1196.29
    5  1.73344  1181.47
    6  1.71466  1194.41
    7  1.72736  1185.62
    8  1.71485  1194.28
    9  1.71440  1194.59
   10  1.72596  1186.59
   11  1.72048  1190.36
   12  1.71808  1192.03
------------------------------------------------------
 Average performance:      1187.50 +- 10.17 GFLOP/s
------------------------------------------------------
* - warm-up, not included in average

I've checked the code myself and added logs and it looks like you are using n_threads = 16 and op_per_it = 8. I was looking into this yesterday before I submitted the issue and assumed op_per_it/frequency was doubled to account for 2 units per core, but didn't realize that was already being accounted for in the threads being doubled. I don't know why I didn't notice this or I did but then quickly forgot, but it looks like that's the issue.

Edit: For hyperthreaded cpus in general, should compute_gflops use max(n_threads, n_cores)?

from peakperf.