Prime Num and Phys, more threads worse results, less threads better results ?

[Markp]

Hi:

I'm new here and I downloaded Passmark PerfomanceTest to compare the results with other benchmark softwares. The results are mostly consistent but the Prime Numbers and Physics. I've a 4 cores, 8 threads Intel CPU with 16GB of RAM. If I set on the configuration the use of all 8 threads I got the best results in everything, but in the Prime Numbers and Physics benchmarks I got the worse ones.

That was strange so I did some tests. I set the configuration to use 4 threads, instead the 8 and I got the worst results in everything but in the Prime Numbers and Physics benchmarks where I got the best ones. So I set the configuration to use 6 threads and I got better results than 4 threads an lower than 8 threads in everything, but in the Prime Numbers and Physics benchmarks it was the opposite.

I added the 3 images with 4, 6 and 8 threads configuration. I didn't do anything else. no BIOS settings, no additional software. I just changed the threads in the Edit/Preferences tab. I did this several times and I got the same results every time.

I guess there's some issues with the multithreading algorithm in those 2 benchmarks with Intel hyperthreading technology, I mean when there aren't physical cores.

regards
Mark

[Simon (PassMark)]

See below:
https://forums.passmark.com/performa...5642#post45642
https://forums.passmark.com/pc-hardw...6662#post46662

Basically for those particular tests which use a lot of memory, you may be hitting a memory bandwidth bottleneck when you starting using more and more threads.
Also for hyperthreading - they are not full CPU cores and don't perform as such. So some algorithms run really well on hyper-threading and give great scaling, others give no benefit, and some have negative scaling.

[Markp]

Hi:

Thanks for your reply. I really appreciate it. I read the links you sent me and yeah, I'm having the same behavior. However, I thought a CPU benchmark software should be aware that SMT/hyperthreading cores aren't full cores. Especially today that most CPUs use that technology. I don't know how useful a CPU test is, if it doesn't scale well. Maybe those 2 test should be added as optional instead mandatory to test the CPU power.

Anyway, I know it's not easy to make these kind of testing software, and this software is what it is.

thanks again for your reply
Mark

[David (PassMark)]

It's the same in real life. Some apps don't scale well and depend on single thread performance. Lots of vendors suggested turning off Hyperthreading (do some Google searches). Some apps benefit and some don't.

Other apps hit the memory bandwidth limit (which is usually a limit based on the memory controller and the cache size, which is in the CPU itself). This is independent of the hyperthreading issue. You can run out of memory bandwidth without using hyperthreading.

If you buy a CPU with more memory bandwidth (and the matching RAM sticks) it will scale better. This is why Intel's next gen CPUs (called Alder Lake) will support faster DDR5, some apps need it. This is also why some CPUs support tri or quad channel RAM and large caches. I am guessing your CPU doesn't however.

I understand the desire to shoot the messenger however.