Prime95 version 29.1

[GP2]

Originally Posted by Prime95

There have been no major complaints with 29.1, so I think it is very safe to use it in a new installation.

Time to update the download page? It's still linking to 28.10

I'm also using 29.2 on many cores for DC work, everything seems fine so far.

[vsuite]

I just upgraded (in place) 3 (soon to be 4) installations of 28.7 build 1 to 29.2 build 4. I hope all goes well.

I assume it is safe to leave 28.10 running on a couple other machines?

I only saw minor (presumably non-statistical) difference between benchmarks of 28.7 and 29.2 on i3-4160 and i5-6400 running multiple workers. But 29.2 seems to be fine tuned to give an extra boost when 1 worker is run on all the threads (which 28.7 does not bring out in the benchmarks).

[Prime95]

Originally Posted by vsuite

I only saw minor (presumably non-statistical) difference between benchmarks of 28.7 and 29.2 on i3-4160 and i5-6400 running multiple workers. But 29.2 seems to be fine tuned to give an extra boost when 1 worker is run on all the threads (which 28.7 does not bring out in the benchmarks).

The FFT code is identical. There may have been a few changes as to which FFT implementation is the default (every FFT size has several possible FFT implementations). The biggest change with 29.2 is that prime95 runs 10 throughput benchmarks of the FFT implementations to select the best FFT implementation for your machine).

[Madpoo]

Originally Posted by GP2

Time to update the download page? It's still linking to 28.10

I'm also using 29.2 on many cores for DC work, everything seems fine so far.

I wondered about that, but I figured maybe 29.x was still considered a work on progress?

Thoughts George? I can update the download page to highlight 29.2 (and generate the SHA for them as well) if you think you're ready for it.

I've updated one of my machines to it... it was a bit of a transition since I had the 28.x config modified every which way with an affinity scramble and all kinds of stuff. I had to make a few tweaks but I imagine the average user never fiddles with it as much. They can probably upgrade and it'll just work.

When I tried it out on my dual 14-core servers, I was curious what kind of default # of workers and cores per worker it would pick, what kind of affinity per worker (only cores from a single CPU, for example).

I wasn't entirely happy with how that worked out. A single worker seemed to randomly pick which cores to use, sometimes across the chip boundary... Maybe I was doing something wrong, but I used the nice, new way of assigning cores to workers (much easier than affinity scramble) and got things the way I like without issue.

As someone who manually submits their results, I was surprised to see the new entries in the results.txt for all of the daily benchmarks. It sounds like those stop once a suitable batch of benches have been done but I only upgraded about a week ago... we'll see.

[mackerel]

Originally Posted by Prime95

The FFT code is identical. There may have been a few changes as to which FFT implementation is the default (every FFT size has several possible FFT implementations). The biggest change with 29.2 is that prime95 runs 10 throughput benchmarks of the FFT implementations to select the best FFT implementation for your machine).

I only use Prime95 as a benchmark and stress test. Does the above affect the manual custom benchmarks? I dread the potential situation where I might run a benchmark on a fresh install, and some later time it might decide to use different FFTs for the same settings on the same system. The way I read it, it sounds like the feature is for real work only, so does that mean the benchmark would only run the default FFT choice for a given size?

I can't recall previously discussion, was there specific consideration given to Skylake-X to account for the very different cache size and speed? I'm about to build such a system and look forward to comparing it against Skylake-S.

[Prime95]

Originally Posted by mackerel

The way I read it, it sounds like the feature is for real work only, so does that mean the benchmark would only run the default FFT choice for a given size?.

In your situation, you'll always run the default FFT implementation.

There are no special settings in prime95 for Skylake-X.

[sanaris]

Hello!

I have seen the option to choose instruction types in local.txt, but can I choose CpuType from that file local.txt? What are possible processor types?

I just know that P95 does not define my L3 size correctly and I'm setting it manually with CpuL3CacheSize (though I don't need it very much for higher efficiency).

[Prime95]

Originally Posted by sanaris

Hello!

I have seen the option to choose instruction types in local.txt, but can I choose CpuType from that file local.txt? What are possible processor types?

I just know that P95 does not define my L3 size correctly and I'm setting it manually with CpuL3CacheSize (though I don't need it very much for higher efficiency).

Don't worry about a mis-identified L3 cache size. It will not affect prime95's FFT selection.