New Multi-Threaded Pi Program - Faster than SuperPi and PiFast

[Movieman]

100M and 250M

[Movieman]

There is a monster in there waiting to be released once we figure out how to OC these.

[tet5uo]

Here's some more runs on the i7 @4.0

[Hoss331]

[Serotoninn]

Ahahaha tet5uo my i7 920 at 4.0 beat out your 965 by 1.739 seconds at 500m
Now go up your bclk by like 1 lol

[tet5uo]

Heheh it's on! :P

Srsly though it's a fast and dirty OC, I just used multi jacked up , my bclk is 133 still

[Serotoninn]

Boo I don't think I can squeeze much more out of my cpu
As things stand it was barely stable at 4Ghz

But that doesn't mean I'm giving up just yet...

[Movieman]

500 Million

[poke349]

Great to see all these benchmarks!

*Just a note, I haven't actually been validating all these screenshots... too much of a pain... and I doubt anyone would try anything stupid this early in the game.

If enough people start to take this program seriously, I'll release the validator to a few "chosen" people. I might even re-enable it in a future public release if it is "hard enough" to reverse engineer.

[tet5uo]

Lol I need more ram... I can't run any bigger tests...

must get another 6gb kit, lol!

[poke349]

Unless you're like me (who can use all that ram and still be begging for more), I wouldn't jump the gun just yet.

My current test build will comfortably do 1 billion digits under 6 GB of ram. (and still have ~700MB left for the OS)

So if you're willing to wait a little bit... (a week - a month?)

2.5 billion digits requires just over 13GB of ram with my current build - not gonna fit under 12GB.


I've had a few people tell me that y-cruncher is a memory hog... but compared to other pi programs, it has one of the best "memory / # of digits" ratios.

SuperPi needs 256MB of ram to do 33m. y-cruncher needs only 200MB. (179MB in my current build)

PiFast 4.3 needs 1.8GB of ram to do 250m at full speed. y-cruncher needs only 1.3GB.

QuickPi 4.5 needs 38.3GB of ram to do 2.5b at full speed. y-cruncher can do it in "only" 15GB (13.2GB in my current build).

[tet5uo]

Heheh It would be totally overkill for this system... I built it mainly for gaming , it also does a-lot of h.264 encoding for me as I'm backing up 1000+ dvd's to a raid array right now.

I'm sure neither of those tasks need more than my 6gb yet, lol.

[spdycpu]

4.3GHz runs (21x205):

Processor(s):   Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz
CPU Frequency:  4,105,000,424 Hz
Thread(s):      8
Digits:         25,000,000
Total Time:     9.48224 seconds
Checksum:       1a607fe8ac76aae51a3a879d7d460c16

Processor(s):   Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz
CPU Frequency:  4,105,000,403 Hz
Thread(s):      8
Digits:         50,000,000
Total Time:     20.553 seconds
Checksum:       8e3711428314f7d606379757726de3f7

Processor(s):   Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz
CPU Frequency:  4,105,000,453 Hz
Thread(s):      8
Digits:         100,000,000
Total Time:     45.081 seconds
Checksum:       7dbb9e8c0d7ee6c80dac0dfc03564ce2

Processor(s):   Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz
CPU Frequency:  4,105,000,424 Hz
Thread(s):      8
Digits:         250,000,000
Total Time:     126.855 seconds
Checksum:       a40277cb8aa985561ec0c0e0d9a06755

[poke349]

Awesome! How much ice did you burn?


I've made some big (more than expected) progress on memory optimizations:




So 1 billion digits can be done with 6GB of ram.

I'm still in the process of massively testing this memory optimization. So I can't release it yet.

[Serotoninn]

Wooo 1billion runs can be made with 6gigs of ram now

The newest version of the program uses 4.79gigs versus the ~5.9 it ate up before. btw this version isn't out yet as OP still needs to do a bunch of testing and I'm helping out atm.

[spdycpu]

About a liter or so. I tried to get 4.4 but by the time I was done testing the ice was gone. I'm surprised at just how much heat that i7 can dump.

Good job on the memory optimizations, looks like I'll be able to do a 500m run with my 3GB of ram. Did the mem optimizations slow the computation time down at all or is it about the same speed (or faster)?

Awesome! How much ice did you burn?


I've made some big (more than expected) progress on memory optimizations:

So 1 billion digits can be done with 6GB of ram.

I'm still in the process of massively testing this memory optimization. So I can't release it yet.

[poke349]

Wow... that's a lot of ice to melt through... lol

Go figure... that's the limit of water cooling? (I've never tried water cooling - always been scared of it...)


The memory optimizations shouldn't affect the speed much. As far as I can tell, it might even be a percent faster too. With the new version, I got a 7280s 10b time - with aero, sidebar, and @ normal priority. It used to take ~7360 +/- 30 or so seconds @ high priority...

Other changes so far:

1. I've added back single-threaded runs for benchmark mode.
2. I've added the program version # to the validation checksum.
3. The Validation.txt file now has the time as well. (It wasn't there before, because it isn't needed to validate the checksum. The time is encrypted into the checksum itself.)

I'm considering enabling the checksum validator in the next public release because I feel that it's "difficult enough" to reverse engineer.

Things that I'm working on:

1. I'm trying to find a better and more reliable way to detect clock frequency.
2. I'll be trying to implement counter-measures for the "slow-motion" cheats that are described here:

http://www.xtremesystems.org/forums/...ad.php?t=46926

All the cheats that are on that thread are obviously a very closely guarded secret. But one of the moderators was kind enough show me one of them so that I can try to counter it in y-cruncher.

3. Lastly, I'm doing a ton of stuff in the custom compute mode that don't apply to benchmarking.

[Serotoninn]

muahaha i got my cpu to 4.1ghz for a bit and did 100m and 500m runs
unfortunately it wasn't stable enough to do a 1b run so i ran that at 4.0ghz lol

[Lightman]

Great little Pi program!
I like it very much!

Here is my entry:
Phenom II 940 @3.7GHz NB2.4GHz Mem@1162MHz 5-5-5-15-26 2T
Vista x64 Business with loads of crap in background!

[wizzard67]

A couple from my test server.

500 million





1 billion



Movieman is right, there's a beast waiting to be unleashed

[spdycpu]

My friend Enrico's Skulltrail @ 4GHz, x64 SSE3 binary:

Processor(s):   Intel(R) Core(TM)2 Extreme CPU X9775 @ 3.20GHz
CPU Frequency:  4,009,000,620 Hz
Thread(s):      8
Digits:         25,000,000
Total Time:     8.46125 seconds
Checksum:       e0c2395f9cedb2cd078cc7e1a753a271

Processor(s):   Intel(R) Core(TM)2 Extreme CPU X9775 @ 3.20GHz
CPU Frequency:  4,009,000,680 Hz
Thread(s):      8
Digits:         50,000,000
Total Time:     17.6719 seconds
Checksum:       2a7b61201ec1133f0bf910993d572693

Processor(s):   Intel(R) Core(TM)2 Extreme CPU X9775 @ 3.20GHz
CPU Frequency:  4,009,000,570 Hz
Thread(s):      8
Digits:         100,000,000
Total Time:     38.1509 seconds
Checksum:       5fab3fc8b275752d0ce14e10b88bc617

Processor(s):   Intel(R) Core(TM)2 Extreme CPU X9775 @ 3.20GHz
CPU Frequency:  4,009,000,550 Hz
Thread(s):      8
Digits:         250,000,000
Total Time:     105.08 seconds
Checksum:       ed26d0a3c5620dd5042fbefa92506d3b

Processor(s):   Intel(R) Core(TM)2 Extreme CPU X9775 @ 3.20GHz
CPU Frequency:  4,009,000,530 Hz
Thread(s):      8
Digits:         500,000,000
Total Time:     230.542 seconds
Checksum:       af68ed11dfd79cd01882c8b6765d693a

Processor(s):   Intel(R) Core(TM)2 Extreme CPU X9775 @ 3.20GHz
CPU Frequency:  4,009,000,640 Hz
Thread(s):      8
Digits:         1,000,000,000
Total Time:     505.422 seconds
Checksum:       c4db0eeb8672859b0943e397d0997215

[poke349]

WoW!!! 8.46125 seconds for 25m... This is starting to look more like the superpi numbers - except 25 x more digits!

That's four #1 spots!!!

Don't know what to say, but I can't wait to see what happens if a big overclocker shows up with some LN2...

*Though I am a little surprised that the 2.8GHz Gainestowns can beat the 4GHz skulltrail at the larger sizes...


Keep these benchmarks coming!!!
Even if you don't have an awesome multi-socket rig post your benchmarks anyway!

If enough workstations dominate the top spots, I'll split the rankings into single-socket and multi-socket divisions.

[wizzard67]

*Though I am a little surprised that the 2.8GHz Gainestowns can beat the 4GHz skulltrail at the larger sizes...

Sorry, forgot to quote specs -

2 x E5560
Intel 5520UR
12GB (6 x 2GB, 6GB per CPU) DDR3 ECC RAM (1066 IIRC)
Various drives

HT and Turbo are both enabled

Running on Windows 2008 x64 Enterprise Server.

[tet5uo]

Just got your PM Poke, I'll check the reg key for you with various multis tonight.

How's that new version coming?

[poke349]

Thanks!

New version looks good so far. The reason why it's taking so long is because the memory optimization affected more than half of the constants in the program.

Specifically: both formulas for Pi, both formulas for Zeta(3), both formulas for Log(2) and Log(10), and the primary formula for Catalan's Constant, were affected by the memory optimization...

So I had to re-test all of these constants for both computation modes at a ton of different sizes... all the way up to 5 billion digits...
So my dual-Harpertown and Serotoninn's i7 920 didn't get much in the way of idle time these past 2 weeks.

Of course only the fast formula for Pi matters for benchmarking , but a lot of Pi-fanatics also play with this program (and it's other features/constants) so I don't want to inadvertently break some of the other constants. (There's this guy in Japan who basically gets brand new 20-grand workstations every year. I don't know what the primary purpose for them is, but he doesn't hesitate to tie them down for months at a time just to crunch digits... His current computer is a Dual Xeon X5470 with 128GB of ram... )


The tests are done now and the program is basically ready to be released. I've already compiled the version that will be released. Right now, I'm doing last minute checks to make sure I didn't break any existing code and to make sure that the improved anti-cheat protection doesn't give any false positives.