Benchmarks

[Pjetrode]

Only a few seconds view on GIMPS main page – Get Started – CPU Benchmarks and I just wonder. The benchmarks there are only for 1 core/1 thread. This kind of CPU (1 core/1 thread) are extremely rare in 2018. The Prime95 application can have the advantage of multi-core and multi-thread CPUs of these days. Therefore I find no reason to publish benchmark only for 1 core/1 thread. You can NOT determine the performance of your CPU simply taking the amount of your CPU cores and looking at the times in that table. Generally it is NOT possible to determine the performance for the specific CPU from this benchmark, because:

(1) The performance progression doesn't need to be linear according the amount of CPU cores, usually it is lower (for instance 6 cores and 5.2-times performance increasing or 12 cores and 9.8-times performance increasing … etc). The variation may be from 5% to 30%, depending on number of cores and CPU architecture. Let us say it is 20%.

(2) Most of mainstream and high-end CPUs have HT-SMT technology, so one physical core can handle two threads. Using double number of threads (compared to number of physical cores on that kind of CPU) may increase the performance in about 20% (for example 6C/6T vs. 6C/12T).

(3) Most of nowadays CPUs have some kind of turbo technology, so they can boost the frequency, when only 1, 2 or a small number of cores/threads is working. When trying to estimate the benchmark for specific CPU (for example on 8 core CPU the time will be 8-times smaller), the benchmark is always optimistic (maybe 10% variation) because all the cores loaded will NEVER run on such boosted frequency like one core does.

When application can use more cores, the only sense is to publish multi-thread benchmarks for CPUs. Considering all three paragraphs, there can be 50% (!!!) variation, when you try to estimate the performance of your CPU only having the information for 1 core performance in Prime95 and when you do not care about core/performance ratio, HT-SMT technology and turbo technology for 1, 2 or a small number of cores/threads working.

[kladner]

Check out the various pages here: www.mersenne.ca/

[VictordeHolland]

Hyper-Threading (SMT) is pretty much useless for Prime95.
With modern multi-core processors, total memory bandwidth is a good proxy of performance.

[Prime95]

Originally Posted by Pjetrode

When application can use more cores, the only sense is to publish multi-thread benchmarks for CPUs.

You are correct. Prime95 was recently changed to send all-cores throughput benchmarks to the server (in addition to single-core timings). Right now the web pages do not have an option to display the throughput benchmark data that has been collected thusfar.