SkylakeX teasers (aka prime95 29.5)

[ATH]

That's more like it: 1.38 ms/iteration at 3072K, should be < 3 ms/ite at 4096K and even at 4608K maybe. That is pretty good, is it still not better than your son's computer?

[petrw1]

Originally Posted by ATH

So on the SPD tab you can see the RAM modules are actually DDR4-2133, the 1066 Mhz is the clock rate and 2133 Mhz is the "effective" speed since they can transfer data twice per clock. You are running Quad channel RAM which is great, the best you can get without buying very expensive Xeon processors.

On the Memory tab it shows that RAM is running overclocked at ~ 1900 Mhz actual clock rate so "effective" rate of 3800 Mhz, and your processor is slightly overclocked to 3.7 Ghz (CPU tab).

I hope you only payed for DDR4-2133 (PC4-17000) modules and not something higher? The question is if it is running stable at 3800 Mhz, you will have to test that with stress test and double checks.


It is strange you are getting bad benchmarks with this setup, it could be throttling because the CPU is getting too hot, try watch the CPU temperature during benchmarks. Also try to benchmark with the 29.4b8 Prime95 (with AVX2 instead of AVX-512), and check the temperature running AVX2.

I don't know what to say about the DDR4-2133. Assuming the techie wasn't lying he read the labels on the RAM and said they are DDR4-3600; my invoice says the same.

He had to OC to 3800 because the RAM is CL18 and the only speeds the BIOS/MB supported with CL18 were 2133, a couple others and 3800.

If he chose a speed (like 3600) that wasn't CL=18 then only 24GB (3 of the 4 sticks)would show.

When he ran MemTest is showed the RAM running at 3598 or something else really close to 3600.

I have a liquid cooler so I hope heat won't be an issue.

Another thread shows much better iteration times with v29.5

[sdbardwick]

Your RAM is DDR4-3600, as verified by the SPD tab in CPU-Z; the XMP profile shows it as 3600 C18 (well, 3596 C18). All RAM is advertised using the XMP figures.

Usually, you just need to tell the BIOS to use the XMP profile (and sometimes manually set the memory voltage to match that in the XMP profile - some motherboards do that automatically, some don't) and you are good to go.

[science_man_88]

Originally Posted by petrw1

If he chose a speed (like 3600) that wasn't CL=18 then only 24GB (3 of the 4 sticks)would show.

really your memory tab says 32 GB quad channel your SPD tab says 8192 Megabytes = 8 Gibibytes

[sdbardwick]

Originally Posted by science_man_88

really your memory tab says 32 GB quad channel your SPD tab says 8192 Megabytes = 8 Gibibytes

Because it is only showing information for the DIMM in slot 1.

[Prime95]

29.5 build 4.

1) Slightly faster small FFTs (under 128K)
2) Reduced round off error for FFT lengths divisible by 7.
3) Proper FFT crossover points.

Not fixed:
1) Switching to AVX FFTs for ultra-tiny tests.
2) The possible stack corruption in erroneously reporting round off errors. Multi-thread crash bugs. These do not seem to be repeatable and could be caused by AVX-512 overheating the CPU. Please continue testing and reporting any issues along these lines.

Linux 64-bit: https://www.dropbox.com/s/67wgupwxlz...64.tar.gz?dl=0
Windows 64-bit (totally untested): https://www.dropbox.com/s/yzfx2jcryf...win64.zip?dl=0

[Chuck]

Just started using the new Windows version. Using 1 worker, 5 threads. Working OK so far. Good improvement in speed.

[Prime95]

29.5 build 5.

1) Fixed AVX FFTs where FFT length was divisible by 7.
2) Fixed several zero padded FFT bugs.
3) Wider CPU dialog box.
4) Correct reporting of Skylake-X L2 cache size.

Not fixed:
1) Switching to AVX FFTs for ultra-tiny tests.
2) The possible stack corruption in erroneously reporting round off errors. Multi-thread crash bugs. These do not seem to be repeatable and could be caused by AVX-512 overheating the CPU. Please continue testing and reporting any issues along these lines.

Linux 64-bit: ftp://mersenne.org/gimps/p95v295b5.linux64.tar.gz
Windows 64-bit (less tested): ftp://mersenne.org/gimps/p95v295b5.win64.zip

[Yuno]

Originally Posted by Prime95

29.5 build 5.

1) Fixed AVX FFTs where FFT length was divisible by 7.
2) Fixed several zero padded FFT bugs.
3) Wider CPU dialog box.
4) Correct reporting of Skylake-X L2 cache size.

Not fixed:
1) Switching to AVX FFTs for ultra-tiny tests.
2) The possible stack corruption in erroneously reporting round off errors. Multi-thread crash bugs. These do not seem to be repeatable and could be caused by AVX-512 overheating the CPU. Please continue testing and reporting any issues along these lines.

Linux 64-bit: ftp://mersenne.org/gimps/p95v295b5.linux64.tar.gz
Windows 64-bit (less tested): ftp://mersenne.org/gimps/p95v295b5.win64.zip

date error on ftp, says p95v295b5.win64.zip 6953 KB 11/29/2017 2:23:00 AM

[Chuck]

This new AVX-512 version still isn't working correctly on my Skylake-X system. When the new build was released, I had an in-process exponent 88042723 running. I exited the program, copied in the new version and libraries, and restarted. Everything seemed fine for several days until it got to the final iterations. The first screen capture shows the results.
I also started getting errors on the next queued exponent 88135807.

I exited the program, deleted the save file, and started 88135807 again. Both runs were using 1 worker, 5 threads. The second screen captures shows the errors.

I erased the save file again and restarted with 1 worker, 1 thread. Still the errors as shown in the third capture. I also uploaded a CPU-Z information image. NOTE when I came back to read this post I see I didn't capture an image with just 1 worker; however, I still got the stream of roundoff errors.

I had to revert back to version 29.4 again. I wish the new AVX version would work for me since it is quite a bit faster.

This is not an issue with the CPU overheating as there is no noticeable difference in the temperature (high 70s).

[Prime95]

Originally Posted by Chuck

This new AVX-512 version still isn't working correctly

Good info. The first 30 and last 30 iterations of an LL test are done differently. Maybe this clue will get me closer to the bug.