Perpetual benchmark thread...

[Viliam Furik]

I've asked paulunderwood to run a few throughput benchmarks on the 3990X. Many thanks for that.
I attached graphs of the results.

The first one is normalized throughput benchmark from size 64K to 19200K for 1, 2, 4, 8, 16, 32, and 64 workers, all using 64 cores in total. Normalization was done by the following formula [normperf = throughput * ln(FFT size) * FFT size / clock speed (4,1 GHz)]. You can see performance drops, which show when the cache runs out.

The second graph is a zoomed-in part of the first one so that the differences can be seen better.

ZIP file contains the values used, as well as original graphs, in an Excel spreadsheet.


Manual data analysis showed some weird behaviour. Due to not having enough time now, they will be mentioned in another post.