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

[poke349]

y-cruncher: Multi-Threaded Pi-Benchmark for Multi-Core Systems... (official thread)

Latest Version: v0.6.5.9442
Now with support for XOP and AVX2. Backward-compatibility is maintained for older processors.

The benchmark is also available for 64-bit Linux.



Download:
http://www.numberworld.org/y-cruncher/

Pi Benchmark Results: Can you do any better?

• 25,000,000 digits
• 50,000,000 digits
• 100,000,000 digits
• 250,000,000 digits
• 500,000,000 digits
• 1,000,000,000 digits
• 2,500,000,000 digits
• 5,000,000,000 digits
• 10,000,000,000 digits
• 25,000,000,000 digits
• 50,000,000,000 digits
• 100,000,000,000 digits
• 250,000,000,000 digits
• 500,000,000,000 digits
• 1,000,000,000,000 digits

-----------------------------------------------------------------------------------------------------

What is y-cruncher?

y-cruncher is (yet another) program that computes Pi to millions of digits.

But unlike SuperPi and PiFast:

• y-cruncher is multi-threaded and is capable of using multiple cores.
• y-cruncher is able to use modern instruction sets such as SSE and AVX.

And unlike HyperPi, y-cruncher uses multiple threads to make a single computation faster. (as opposed to running the same computation multiple times)

Because y-cruncher is multi-threaded and uses up-to-date processor features, it can reach speeds and sizes that are unobtainable with older apps:

• A stock AMD FX-8350 can do 32M digits in 9.737 seconds using y-cruncher. SuperPi requires upwards of 20 minutes.
• A Core i7 4770K @ 4.0 GHz can do 1,000,000,000 digits in 206 seconds.
• The same Core i7 equipped with 32GB ram and 8 hard drives can do 100 billion digits in 23 hours.
• A dual Xeon E5-2690W @ 3.5 GHz with 256GB ram and 22 hard drives can do 1 trillion digits in 83 hours.

Likewise, y-cruncher is capable of setting world records for the most digits ever computed:

• December 2013 - 12.1 trillion digits
• October 2011 - 10 trillion digits
• August 2010 - 5 trillion digits

[poke349]

2.4 GHz Kentsfield

32m

SuperPi: 1289.29 seconds
PiFast 4.3: 122.41 seconds
QuickPi 4.5 (x86): 58.84 seconds
y-cruncher (x64): 30.357 seconds

64-bit QuickPi surprisingly ran slower than 32-bit.

(click to enlarge)

[Serotoninn]

Hi guys I ran a comparison too computing pi to 32m with a core i7 920 @ 3.2ghz

superpi: 660.614s
pifast 4.3: 77.50s
quickpi x64: 32.95s
y-cruncher x64 16.692s

and not shown on the screenie, quickpi 32bit took 34.35s



y-cruncher = full of win

[poke349]

wth... XS reverted...

Anyways... New version is out! This one has benchmark validation and anti-cheat protection!!!

[ChaosMinionX]

Cool


Checking it out now

[spdycpu]

wth... XS reverted...

Anyways... New version is out! This one has benchmark validation and anti-cheat protection!!!

Wish I had saved all the results posted in this thread.

I'm curious again about a few things:
Do you think you'll maintain a result list?
In the Validation.txt output could you include the time?
Could you bring back the ability to change the number of running threads?
Maybe add an efficiency number (cycles per digit for example)?

For example my work system is an E7200 (Penryn dual core 2.53GHz), here are the results with the x86 SSE3 binary:

Processor(s):   Intel(R) Core(TM)2 Duo CPU E7200 @ 2.53GHz
CPU Frequency:  2,527,040,841 Hz
Thread(s):      2
Digits:         25,000,000
Total Time:     56.4712 seconds
Checksum:       46d5028e60d7bba608077b159b3d8a4b

If we do...
( (CPU Frequency * number of actual cores) * seconds) / digits calculated
we get 11416.4 cycles per digit. Would it be something worth adding you think? The CPD timing would likely only be valid for a set number of digits, but, if people like to test one size (100,000,000 for example) then it could be useful for comparison and tweaking.

What do you think?

[poke349]

About half the old posts are still in the Google cache. Should we grab them and re-post?


Result List:
I'll start by listing them all on this thread (on my first post.)
I'll keep track of the top few times on each size. But I won't split them into different binaries (x86 vs. x64) because that would be too much. (Yes, let 64-bit rule!!!)

Change Threads:
Yes I can bring that back. But it can only have 2 modes: 1 thread, or threads >= cores (first power of 2 that's >= # of cores).
This is because of the way the program uses more threads than selected, so "extra cores" will skew results. There's no way I can enforce CPU affinity.

Efficiency Number:
Yes I can do that. But keep in mind that it won't scale across different sizes and # of cores.
Computing Pi isn't linear. It's roughly O( n*log(n)^3 ). So larger computations will have lower efficiency.
Also, the program doesn't quite scale linearly with the # of cores. So multi-threaded benchmarks will have lower efficiency.



Aside from that:

The next version, will use less memory. 1 billion digits will only require 5.25GB as opposed to 5.98GB of ram in the current release. Which means that it'll be possible for the "standard" Core i7 with 6GB of ram to bench 1 billion digits.

For the "very large" end, 10 billion digits will require 52.9 GB of ram (in the new version). By the looks of it, I "might" be able to squeeze that down to 46GB - which would make it possible for a fully-loaded 12-slot Gainestown server to do. And THAT would be interesting...

Anyways... Post your benchmarks everyone!!!

[moogle]

Anyways... Post your benchmarks everyone!!!

Nice to see a new Pi software finally made

[WhiteFireDragon]

how do you cheat in benching with superpi?

[poke349]

Updated the list of benchmarks with old results (from before the XS revert) and with results from other forums.


Stuff like tampering with the system clock, output tampering... is standard for cheating.

I've made sure none of that will work on y-cruncher.



Also, if everyone copy and pastes the validation text into your posts too. It would make it easier for me to validate.

[tet5uo]

Cool program!

Here's a quick test on my new rig.



It's core i7 @ 4.0gHz with 1600mHz DDR3 (QPI is only stock , oc'ed for 24/7 use with multiplier since It's a 965.)

[poke349]

Interesting... Your benchmark is valid (the checksum is good). But the CPU frequency is wrong... And by the very impressive timing, I can tell it really IS at 4GHz.

I swear there's something wrong with the rdtsc instruction... Because it doesn't detect turbo boost from 3.2 to 3.4 on my friend's computer. (it reports it at 3.2 even with turbo kicked on)

Maybe I'm not using it right.

Problem is: I actually don't know how to fix it. Anyone, by any random chance, know how to do this properly?

[tet5uo]

I'm awful at programming, so not much help here

Is there any way to get this data from the BIOS directly?

I noticed that alot of apps misread my CPU speed, the only thing that seems to get it right is CPU-Z or the EVGA version.

[poke349]

Okay, then I guess I'm not alone...

Would you mind if you try some of the smaller benchmarks?

At 4Ghz, you might be able to beat my dual-Harpertown rig at 25 and 50 mil.

Someone with a 920 OC'ed to 4.1 had a faster 25m and 50m time than my dual-Harpertown rig. But it's gone from the revert and it isn't in the Google cache, so I can't recover it.

[tet5uo]

Sure, I'll run a couple more tests and edit them into this for you.

Thanks for making this program I'm sure we all love to play with new benches

[spdycpu]

I did some runs at 4.1 (20x205) but I'm messing around now at 21x200 w/ turbo mode. As mentioned before it doesn't seem to detect that I'm running at 4.2GHz. I tried 21x205 for 4.3GHz but unfortunately my 2lbs of ice magically disappeared after about 2 seconds. I only managed a single run of 25m:

4.3GHz @ 21x205

Processor(s):   Intel(R) Core(TM) i7 CPU 920 @ 2.67GHz
CPU Frequency:  4,105,000,554 Hz
Thread(s):      8
Digits:         25,000,000
Total Time:     9.56572 seconds
Checksum:       3c277379f4ec9d5316e5b430710d0cbf

Interestingly enough 4.2GHz is stable in Prime95 and Linx, 4.3 just gives me fits however. 4.1GHz seems to run slightly faster in y-cruncher than 4.2GHz, I'd imagine due to the faster L3 and memory. More ice (MUCH more) is in the freezer, I'll try some 4.3GHz runs at 50, 100, 250 tomorrow.

Okay, then I guess I'm not alone...

Would you mind if you try some of the smaller benchmarks?

At 4Ghz, you might be able to beat my dual-Harpertown rig at 25 and 50 mil.

Someone with a 920 OC'ed to 4.1 had a faster 25m and 50m time than my dual-Harpertown rig. But it's gone from the revert and it isn't in the Google cache, so I can't recover it.

[poke349]

Awesomeness... I'll be looking forward to those...

Finally I can kick myself off the top spot for one of the sizes.

I expect to hold the top spot only for the larger sizes... simply because no (overclockable) machine has the memory for it...

I'd love to see someone with dual W5580s and 72GB of ram show up and clean sweep these benchmarks...

Yes, I'll need to put the single-thread mode back in. Otherwise, y-cruncher is going to turn into a who-has-the-most-money server competition... (especially so with the ram requirement)
I don't exactly want to see the top spots being dominated by quad socket Dunningtons of some sort or whatever the new one coming out is... (As is the wprime benchmarks on hwbot...)

[Movieman]

Awesomeness... I'll be looking forward to those...

Finally I can kick myself off the top spot for one of the sizes.

I expect to hold the top spot only for the larger sizes... simply because no (overclockable) machine has the memory for it...

I'd love to see someone with dual W5580s and 72GB of ram show up and clean sweep these benchmarks...

Yes, I'll need to put the single-thread mode back in. Otherwise, y-cruncher is going to turn into a who-has-the-most-money server competition... (especially so with the ram requirement)
I don't exactly want to see the top spots being dominated by quad socket Dunningtons of some sort or whatever the new one coming out is... (As is the wprime benchmarks on hwbot...)

Will dual X5570's do for now?

[poke349]

Are you SERIOUS?!?!?!


Let it roll!!!

You should be able to boot me off of everything you have enough ram for...

[Movieman]

Are you SERIOUS?!?!?!


Let it roll!!!

You should be able to boot me off of everything you have enough ram for...

Just 6 gig..YGPM

[[XC] Hicks121]

Just 6 gig..YGPM

aww come on Dave, lets see it.

[Movieman]

This any good?

[poke349]

Oh wait, use the newest version, v0.3.1.

It has a benchmark mode fully equipped with my "lame" attempt at validation.

Also see if you guys can cheat the times too... I'm no expert in cheating benchmarks so I wouldn't know how to write a decent anti-cheat protection.

[Movieman]

Here's 25M and 50M with the latest version of the app:

[poke349]

Updated Your 25m re-run is slower probably because of normal variation. +/- .1 seconds is very typical for so many threads. (I see it myself on my Harpertowns...)

You should try some of the larger sizes... The program doesn't scale well with multi-threading for small computations.

You'll need to super-size it to at least 250 million before it can bring out the power of those 16-virtual cores.