Prime95 version 26.2

[Prime95]

Originally Posted by delta_t

This is the entire prime.log from when I tried to run 26.2:.

Can I trouble you to try 26.2 again? This time go to the Test/Primenet... dialog box. Go to Connection... Select Verbose debug output. Send the prime.log file after it acts up again.

Thanks.

[delta_t]

Originally Posted by Prime95

Can I trouble you to try 26.2 again? This time go to the Test/Primenet... dialog box. Go to Connection... Select Verbose debug output. Send the prime.log file after it acts up again.

Thanks.

It may be another day or two before I get back to that machine, but I'll do as you indicated and get you the new prime.log.
Thanks

[joblack]

What per cent speed impact does the activation of the Round off checking and SUMS(input) error (both or only one of them) have?

[Prime95]

Originally Posted by joblack

What per cent speed impact does the activation of the Round off checking and SUMS(input) error (both or only one of them) have?

Probably 2-4%. Try it.

[joblack]

Originally Posted by Prime95

Probably 2-4%. Try it.

I'm puzzled. Didn't you suggest that a round error wouldn't almost be possible anyway?

[Rhyled]

I ran the benchmark option of a freshly loaded p95v262.zip and noticed a couple of anamolies on my Core i7-920 @3.71 GHz system

In the worker window, the benchmark spams a bit in v26.2:

[Sep 26 19:22] Timing 10 iterations at 5120K FFT length. Best time: 20.389 ms., avg time: 20.724 ms.
[Sep 26 19:22] Setting affinity to run helper thread 1 on logical CPU #1
[Sep 26 19:22] Setting affinity to run helper thread 3 on logical CPU #3
[Sep 26 19:22] Setting affinity to run helper thread 2 on logical CPU #2
[Sep 26 19:22] Setting affinity to run helper thread 1 on logical CPU #1
[Sep 26 19:22] Setting affinity to run helper thread 2 on logical CPU #2
[Sep 26 19:22] Setting affinity to run helper thread 3 on logical CPU #3
[Sep 26 19:22] Setting affinity to run helper thread 1 on logical CPU #1
[Sep 26 19:22] Setting affinity to run helper thread 3 on logical CPU #3
[Sep 26 19:22] Setting affinity to run helper thread 2 on logical CPU #2
[Sep 26 19:22] Setting affinity to run helper thread 1 on logical CPU #1
[Sep 26 19:22] Setting affinity to run helper thread 2 on logical CPU #2
[Sep 26 19:22] Setting affinity to run helper thread 3 on logical CPU #3
[Sep 26 19:22] Timing 10 iterations at 6144K FFT length. Best time: 25.928 ms., avg time: 26.193 ms.

Also, the Trial Factoring benchmark isn't very believable. Note the sudden jump between 62 & 63 bits.
Sep 26 19:22] Timing trial factoring of M35000011 with 58 bit length factors. Best time: 2.620 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 59 bit length factors. Best time: 2.623 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 60 bit length factors. Best time: 2.630 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 61 bit length factors. Best time: 2.621 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 62 bit length factors. Best time: 2.642 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 63 bit length factors. Best time: 4.423 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 64 bit length factors. Best time: 4.414 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 65 bit length factors. Best time: 4.049 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 66 bit length factors. Best time: 4.023 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 67 bit length factors. Best time: 4.019 ms.

I love the ~15% speed up in 2560K FFTs compared to v25.11 ( 32ms vs 38ms)

I find it hard to believe that trial factoring is actually slower in 26.2, but that's what the benchmark says (up to 10%, depending on the factor size).

Here are my result files from v25.11 and v26.2 benchmarking, plus a copy of Worker window 1 from v26: Results & Worker Window.zip

[petrw1]

Originally Posted by Prime95

Yes, it is OK to change FFT sizes mid-LL test.

The CPU credit should drop proportionally. However, I haven't checked that this Primenet code is working properly. The server has a table with iteration times for 2048K and 2560K FFT. It should interpolate to get the 2240K FFT iteration time.

Well the assignments completed and it appears to have given me the same credit V25 would have given me for a 2560FFT.

[liqi]

Originally Posted by Rhyled

Also, the Trial Factoring benchmark isn't very believable. Note the sudden jump between 62 & 63 bits.
Sep 26 19:22] Timing trial factoring of M35000011 with 58 bit length factors. Best time: 2.620 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 59 bit length factors. Best time: 2.623 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 60 bit length factors. Best time: 2.630 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 61 bit length factors. Best time: 2.621 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 62 bit length factors. Best time: 2.642 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 63 bit length factors. Best time: 4.423 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 64 bit length factors. Best time: 4.414 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 65 bit length factors. Best time: 4.049 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 66 bit length factors. Best time: 4.023 ms.
[Sep 26 19:22] Timing trial factoring of M35000011 with 67 bit length factors. Best time: 4.019 ms.

I love the ~15% speed up in 2560K FFTs compared to v25.11 ( 32ms vs 38ms)

I find it hard to believe that trial factoring is actually slower in 26.2, but that's what the benchmark says (up to 10%, depending on the factor size).

trial factoring timing is quite different between 32-bit version and 64-bit verison.
64-bit is much faster than 32-bit version.

please benchmark Prime95 64-bit version 26.2 and compare with Prime95 64-bit version 25.11, and check the results.

[liqi]

Originally Posted by liqi

trial factoring timing is quite different between 32-bit version and 64-bit verison.
64-bit is much faster than 32-bit version.

please benchmark Prime95 64-bit version 26.2 and compare with Prime95 64-bit version 25.11, and check the results.

Use this one:
Windows 64-bit: ftp://mersenne.org/gimps/p64v262.zip

[henryzz]

This is a gwnum problem not cllr. For some reason cllr is using a pentium 4 fft method on my Q6600. Why not the core 2 one? How does it select which to use?

Starting Lucas Lehmer Riesel prime test of 595*2^910447-1
Using Pentium4 type-3 FFT length 64K, Pass1=256, Pass2=256

[Prime95]

Originally Posted by henryzz

This is a gwnum problem not cllr. For some reason cllr is using a pentium 4 fft method on my Q6600. Why not the core 2 one? How does it select which to use?

Starting Lucas Lehmer Riesel prime test of 595*2^910447-1
Using Pentium4 type-3 FFT length 64K, Pass1=256, Pass2=256

This is normal, albeit confusing. Gwnum consists of building block macros that are optimized for each architecture. These are then put together along with loops and prefetching code to make the FFT.

1) You are probably using a 32-bit executable. The difference between Pentium 4 optimized building blocks and Core 2 optimized building blocks is minimal -- no extra registers available.
2) The Pentium 4 prefetch instruction loads 128 bytes, a Core 2 prefetches 64 bytes. Thus a Core 2 optimized FFT has twice as many prefetch instructions.
3) A Core 2 chip has lots of L2 cache. A 64K FFT probably keeps most of its data in cache, making prefetch instructions of little to no value. Thus, a Pentium4 optimized FFT might be a little faster because it wastes less time executing useless prefetch instructions.

Anyhow, the FFT that is selected came from me doing actual timings of Pentium-4 and Core2 optmized FFTs. The Pentium4 FFT was a hair faster.

Perhaps I should change the FFT description to "Using Pentium4-optimized-even-though-this-is-a-Core2-CPU type-3 FFT"