Prime95 version 27.6

[Prime95]

Originally Posted by LaurV

@George: just to make it clear, for me and others, "official" means that core2 workers should update too? (gains? penalties? would they feel the speedup, or is only for heavy artillery like AVX?) or they should better live with v26?

Core2 shouldn't be any faster or slower. Upgrading to v27.6 is optional. Only Intel AVX CPUs will enjoy the big gains. "Official" means that it will be the version available at http://mersenne.org/freesoft.htm

[pepi37]

Originally Posted by Prime95

Bulldozer uses the SSE2 FFTs. AVX FFTs on Bulldozer are slower than the SSE2 FFTs.

So how to determine that AVX is enabled and runing in BD CPU-s?
I try this version on BD CPU, and as you say it uses SSE2. On my Intel I5 it says that use AVX...

[BigBrother]

Originally Posted by Prime95

Bulldozer uses the SSE2 FFTs. AVX FFTs on Bulldozer are slower than the SSE2 FFTs.

Yet another Bulldozer fail... Will AMD ever make anything competitive again?

[Dubslow]

Originally Posted by pepi37

So how to determine that AVX is enabled and runing in BD CPU-s?
I try this version on BD CPU, and as you say it uses SSE2. On my Intel I5 it says that use AVX...

Prime95 is able to detect which CPU it's running on, and choose the best FFTs; when it detects your BD, it recognizes that SSE2 is faster and chooses that; if it recognizes Intel SB or IB, then it knows that AVX is faster and chooses that.

Put another way: AVX FFTs do work on BD, but they're just too slow.

Originally Posted by BigBrother

Yet another Bulldozer fail... Will AMD ever make anything competitive again?

Keep in mind that people were saying the same around the time of Athlon/Pentium, and look where we are now. Don't be surprised if we've flip flopped again in another 5-10 years.

[Prime95]

Originally Posted by dabaichi

I just did some benchmarks on both version 26.6 and 27.6. It turned out that the result puzzled me a lot. On my laptop (running 64-bit Windows 7), the benchmark results are normal. But on my desktop (running 32-bit Windows XP), it seems that trial factoring significantly slows down on version 27.6.

Fix uploaded for Windows and Mac OS X. Other OSes to follow later.

[BigBrother]

Originally Posted by Dubslow

Put another way: AVX FFTs do work on BD, but they're just too slow. Keep in mind that people were saying the same around the time of Athlon/Pentium, and look where we are now. Don't be surprised if we've flip flopped again in another 5-10 years.

My crystal ball is temporarily out of order, you might be right.
But it seems that AMD will be churning out variations of Bulldozer for the next couple of years (Piledriver, Steamroller, Excavator). Unless AMD makes major improvements to Bulldozer in both performance and energy consumption, Intel will easily beat AMD, at least in the HPC/GIMPS department (and what else should you be using a computer for?).

[harlee]

Originally Posted by Prime95

Core2 shouldn't be any faster or slower.

What about Pentium 4s with 512K L2 cache or less? On my system, the 27.6 benchmarks seems to be slower then 26.6 and P-1 factoring seems to be slower. I'm planning on doing more testing this weekend to get more data using different FFT sizes. Also, I've noticed that the type of FFT being used is now being reported differently (26.6 - Using Pentium4 type-3 FFT length vs 27.6 - Using Pentium4 FFT length) and cpu type is missing on my MacOSX Core2 Duo (using FFT length).

[Prime95]

Originally Posted by harlee

What about Pentium 4s with 512K L2 cache or less?

Timings should be virtually unchanged.

Also, I've noticed that the type of FFT being used is now being reported differently (26.6 - Using Pentium4 type-3 FFT length vs 27.6 - Using Pentium4 FFT length) and cpu type is missing on my MacOSX Core2 Duo (using FFT length).

Yeah, I changed the text output a little. type-3 FFT, meaningless text to everyone except me, is by far the most common two-pass FFT. Consequently, I only output the FFT type if it isn't type-3.

[BrianOC]

Hi. My problem:

27.6: http://i50.tinypic.com/345fp1j.png

27.4:http://i45.tinypic.com/2ylps9s.png

Bios reset stock all.
i7 2600, 8GB RAM, GTX 480

This bug? Sorry my english bad. Thanks!

[LaurV]

No bug.
Use the last version 27.6
Use only 4 workers (i7-2600 has only 4 physical cores, the rest are logical - hyperthreaded - cores, using 8 cores produce additional heat, without additional performance, well, somehow).
Also, the versions with AVX will stress that particular CPU more then the versions with SSE (AVX will use the CPU more intensely). I have 2600k and I can run 4 workers, each with helper HT on v26.6 at 4.5GHz, but when I run v27.6 I have either to reduce the clock or either kill the helpers, otherwise I get blue screens. Of course, 27.6 gets better performance, despite of reduced clock and missing helper threads.

[Stef42]

The core voltage seems very low, it should be more like 1.2+ instead of 1.1V
Even so, because you're running stock it should be working normally. Have you tried other stress programs?