Prime95 version 27.1 early preview, not-even-close-to-beta release

Prime95 · 2011-12-05, 23:38

A very early pre-beta prime95 version 27.2 is available. It has support for most, but not all AVX FFT lengths. I have not done any of the 64-bit optimizations.

I've been using this version on my Sandy Bridge for the last few days without incident. You are free to do so too. I'm fairly confident it will produce valid results as it has passed some torture testing and QA runs. If you do not have a Sandy Bridge CPU (or maybe a Bulldozer) there is absolutely no reason to download this version. In fact it would be dangerous to do so as I have not tested the SSE2 or x87 FFTs to see if I've broken something!

Download links:
Windows 32-bit: ftp://mersenne.org/gimps/p95v272.zip
Linux 32-bit: ftp://mersenne.org/gimps/mprime272.tar.gzSource code: ftp://mersenne.org/gimps/source272.zip

I'd be happy to hear of bug reports. I'd be curious about Bulldozer benchmarks. A few double-check LL tests would improve my confidence in the FFT code.

Next up for me is 64-bit optimizations. These things take time. Please be patient.

kdgehman · 2011-12-06, 00:28

This is great!

I started a 46M double check and a 53M double check.

Primeinator · 2011-12-06, 00:46

Your dedication is truly inspiring, Mr. Woltman!

p.s. Is it not about time for you to go on vacation again?

Dubslow · 2011-12-06, 01:29

Woohoo! How necessary are testers? And can it tell whether a longer AVX will be faster than a shorter non-AVX? I'd rather only run a well tested version, but if you need support...
On further thought, I have 5 LL's to finish, then a bunch of DC's in queue. If you don't need testers right now, then I'll switch then.

bcp19 · 2011-12-06, 03:18

Is this part of the reason the worktodo is starting to have things like FFT2=1344K and FFT2=2400K in them?

Batalov · 2011-12-06, 03:20

No FFT2=xxx modifications are old, ...but the FFT sizes may be new!

Dubslow · 2011-12-06, 04:20

Just how much of a performance gain have you gotten George? 5%? 15%? (20%?

)

Brain · 2011-12-06, 21:21

Just 2 stripped benchmarks, 26.4 vs 27.1, running with turbo @2.7 GHz:
--> about +18% to +25%.
26.4:

Code:

Intel(R) Core(TM) i7-2630QM CPU @ 2.00GHz
CPU speed: 1995.49 MHz, 4 hyperthreaded cores
CPU features: RDTSC, CMOV, Prefetch, MMX, SSE, SSE2, SSE4
L1 cache size: unknown
L2 cache size: 256 KB, L3 cache size: 6 MB
L1 cache line size: unknown
L2 cache line size: 64 bytes
TLBS: 64
Prime95 64-bit version 26.4, RdtscTiming=1
Best time for 768K FFT length: 8.876 ms., avg: 9.077 ms.
Best time for 896K FFT length: 10.881 ms., avg: 11.136 ms.
Best time for 1024K FFT length: 11.874 ms., avg: 12.118 ms.
Best time for 1280K FFT length: 15.445 ms., avg: 15.701 ms.
Best time for 1536K FFT length: 18.722 ms., avg: 19.108 ms.
Best time for 1792K FFT length: 22.731 ms., avg: 23.195 ms.
Best time for 2048K FFT length: 25.190 ms., avg: 25.595 ms.
Best time for 2560K FFT length: 32.092 ms., avg: 32.462 ms.
Best time for 3072K FFT length: 39.602 ms., avg: 39.933 ms.
Best time for 3584K FFT length: 47.419 ms., avg: 48.026 ms.
Best time for 4096K FFT length: 52.857 ms., avg: 53.264 ms.
Best time for 5120K FFT length: 68.475 ms., avg: 69.319 ms.
Best time for 6144K FFT length: 85.146 ms., avg: 85.692 ms.
Best time for 7168K FFT length: 104.642 ms., avg: 105.383 ms.
Best time for 8192K FFT length: 114.059 ms., avg: 114.784 ms.

27.1:

Code:

Intel(R) Core(TM) i7-2630QM CPU @ 2.00GHz
CPU speed: 2000.00 MHz, 4 hyperthreaded cores
CPU features: Prefetch, MMX, SSE, SSE2, SSE4, AVX
L1 cache size: 32 KB
L2 cache size: 256 KB, L3 cache size: 6 MB
L1 cache line size: 64 bytes
L2 cache line size: 64 bytes
TLBS: 64
Prime95 32-bit version 27.1, RdtscTiming=1
Best time for 768K FFT length: 7.048 ms., avg: 7.141 ms.
Best time for 1024K FFT length: 9.686 ms., avg: 10.667 ms.
Best time for 1280K FFT length: 12.758 ms., avg: 12.819 ms.
Best time for 1536K FFT length: 15.474 ms., avg: 15.642 ms.
Best time for 2048K FFT length: 20.529 ms., avg: 20.680 ms.
Best time for 2560K FFT length: 26.885 ms., avg: 27.446 ms.
Best time for 3072K FFT length: 32.224 ms., avg: 32.752 ms.
Best time for 4096K FFT length: 45.063 ms., avg: 45.368 ms.
Best time for 5120K FFT length: 57.751 ms., avg: 59.270 ms.
Best time for 6144K FFT length: 69.501 ms., avg: 70.132 ms.
Best time for 8192K FFT length: 96.527 ms., avg: 97.154 ms.

Ralf Recker · 2011-12-06, 21:38

Quote:

Originally Posted by Prime95

If there is demand, I could also create a 32-bit mprime (would that run on a 64-bit Linux install)?

Yes.

Brain · 2011-12-06, 21:50

Expected? A problem? Just replaced 64bit 26.6 with 32bit 27.1...
I know it's no serious problem... I don't know which input parameters go into the identity algorithm... (although I can imagine some)

Code:

[Comm thread Dec 6 22:44] Updating computer information on the server
[Comm thread Dec 6 22:44] PrimeNet error 33: CPU identity mismatch
[Comm thread Dec 6 22:44] CPU identity mismatch: g=6fc985a478b98b3d0d97f721b3e5c8f5 hg=c38f563e9106c58dd3581a7d503d856d wg=cf45fda07e3dcea005efae6e48d62c90
[Comm thread Dec 6 22:44] Updating computer information on the server
[Comm thread Dec 6 22:44] Exchanging program options with server
[Comm thread Dec 6 22:44] Sending expected completion date for M26451023: Dec 15 2011
[Comm thread Dec 6 22:44] Done communicating with server.

Dubslow · 2011-12-07, 01:23

Actually I just realized I'll need MPrime to do my testing... but not until around Christmas, when my last current LL finishes.

2011-12-05, 23:38	#1
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 11×719 Posts	Prime95 version 27.1 early preview, not-even-close-to-beta release A very early pre-beta prime95 version 27.2 is available. It has support for most, but not all AVX FFT lengths. I have not done any of the 64-bit optimizations. I've been using this version on my Sandy Bridge for the last few days without incident. You are free to do so too. I'm fairly confident it will produce valid results as it has passed some torture testing and QA runs. If you do not have a Sandy Bridge CPU (or maybe a Bulldozer) there is absolutely no reason to download this version. In fact it would be dangerous to do so as I have not tested the SSE2 or x87 FFTs to see if I've broken something! Download links: Windows 32-bit: ftp://mersenne.org/gimps/p95v272.zip Linux 32-bit: ftp://mersenne.org/gimps/mprime272.tar.gzSource code: ftp://mersenne.org/gimps/source272.zip I'd be happy to hear of bug reports. I'd be curious about Bulldozer benchmarks. A few double-check LL tests would improve my confidence in the FFT code. Next up for me is 64-bit optimizations. These things take time. Please be patient. Last fiddled with by Prime95 on 2011-12-20 at 01:58

2011-12-06, 00:46	#3
Primeinator "Kyle" Feb 2005 Somewhere near M52.. 3×5×61 Posts	Your dedication is truly inspiring, Mr. Woltman! p.s. Is it not about time for you to go on vacation again? Last fiddled with by Primeinator on 2011-12-06 at 00:47

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Prime95 beta version 28.4	Prime95	Software	20	2014-03-02 02:51
Prime95 beta version 28.3	Prime95	Software	68	2014-02-23 05:42
Beta version 24.12 available	Prime95	Software	33	2005-06-14 13:19
Early Beta of version 24.11	Prime95	Software	113	2005-05-24 17:05
Announcing the release of Triance OS 1.0-BETA	zariss	Linux	1	2004-08-23 21:42

2011-12-06, 00:28	#2
kdgehman Feb 2011 2²·3 Posts	This is great! I started a 46M double check and a 53M double check.

2011-12-06, 01:29	#4
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	Woohoo! How necessary are testers? And can it tell whether a longer AVX will be faster than a shorter non-AVX? I'd rather only run a well tested version, but if you need support... On further thought, I have 5 LL's to finish, then a bunch of DC's in queue. If you don't need testers right now, then I'll switch then.

2011-12-06, 03:18	#5
bcp19 Oct 2011 7·97 Posts	Is this part of the reason the worktodo is starting to have things like FFT2=1344K and FFT2=2400K in them?

2011-12-06, 03:20	#6
Batalov "Serge" Mar 2008 Phi(4,2^7658614+1)/2 2685₁₆ Posts	No FFT2=xxx modifications are old, ...but the FFT sizes may be new!

2011-12-06, 04:20	#7
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	Just how much of a performance gain have you gotten George? 5%? 15%? (20%? )

2011-12-07, 01:23	#11
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 1110000110101₂ Posts	Actually I just realized I'll need MPrime to do my testing... but not until around Christmas, when my last current LL finishes.