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

[poke349]

Ran the batch benches with the new 0.4.4. Here are my results:

25000000 8.074 7573850b0bc23f8e1bee78aaa4354166
50000000 17.922 864abf38181e874d028bf4272c443e26
100000000 37.940 48a784fc35471ec408603601ec00df76
250000000 107.450 647c8e7c845bbb4c1b732a2d71c57923
500000000 234.795 1174e0882e3932251d65f55e6d8c44dc
1000000000 519.458 677f930fa3dcd53514e3e5c8dee7ae59
2500000000 2433.400 8f48e6e52fdfca981d82a0005220b7b7


Guess there's no page-trashing this time.

Cool

btw, you need the rest of the stuff at the top of file to complete the batch validation. But I'll accept these anyway.

(Not like I've been validating any benchmarks anyway... )


EDIT:
What frequency is this at? (Before and after turbo boost)
And I'm assuming Core i7 920 and 12GB of ram?

[Talonman]

My Q6600 at a mild 3.6GHz: (96.5633% Multi-Core efficiency.)

[2good4you]

Hot application!!!

But, why does the K10 CPU's run at SSE3 instructionset and not the supported SSE4? The Intel CPU's run on it's supported SSE4.1. Thanks for the program and an answer would been appreciated. Happy new year from Sweden also.

[poke349]

But, why does the K10 CPU's run at SSE3 instructionset and not the supported SSE4? The Intel CPU's run on it's supported SSE4.1. Thanks for the program and an answer would been appreciated. Happy new year from Sweden also.

SSE4.1 and SSE4a are different instruction sets.

Intel has SSE4.1.
AMD has SSE4a.

In my opinion, there's nothing in the SSE4a instruction set that is useful for this program.

EDIT:
And even it did, I don't have access to a K10 machine with enough ram to properly test it. (Since an SSE4a version won't run on my Xeon workstation.)

EDIT 2:
The AMD optimized (Kasumi) binary was also tested on my Xeon workstation for correctness. (along with all the other x64 binaries)
For correctness testing, it doesn't matter that I'm using an Intel machine to test an AMD binary.
Only the performance tuning had to be done on an AMD machine - which was a Phenom II X3 unlocked to 4 cores.

And yes, Happy New Year!!!

[poke349]

Just tried a run on a Core i7-720QM laptop...

My recommendation... not a very good idea...

Look at these temps: (click to enlarge)



4-core load can do +1 turbo... No chance under these conditions...

[Chumbucket843]

so apparently you have been beat, not in speed though.
http://bellard.org/pi/pi2700e9/announce.html

[poke349]

so apparently you have been beat, not in speed though.
http://bellard.org/pi/pi2700e9/announce.html

I was never beat. Since I never had the Pi record.

jk...

*btw, I knew long before you posted this.

This is actually a lot more impressive than the previous supercomputer record...
Especially the fact that he did it using only 6 GB of ram.

He made no mention of how fast it is for ram-only computations.
Given that disk IO is a huge factor, there's no way to guesstimate how fast it is from the data he gives.

I also can't comment on how fast his program for the disk-sizes since I haven't finished writing my own "Advanced Swap" mode yet.
Though it sure is a hella lot faster than QuickPi 4.5.


His pdf about the algorithms he used is pretty interesting...
There's a number of them I've never even heard of... not a surprise given that he's a world-renowned mathematician. (unlike me... )



So we can't say anything about speed until we can benchmark his program.

Judging by his paper, I can tell he has taken optimization very far... a lot more than me.
So I wouldn't be at all surprised if it is faster than y-cruncher.

[StevensDE]

Here are my new results with a Dell Precision T7400 with 2x Intel Xeon X5450 Harpertown (Revision E0)

[botld92z]

i7-920 @ 4.0ghz, no turbo boost
6gb OCZ Gold "low voltage" memory @ 1600mhz
Asus P6X58D Premium

[poke349]

Here are my new results with a Dell Precision T7400 with 2x Intel Xeon X5450 Harpertown (Revision E0)

Nice! A new machine?

i7-920 @ 4.0ghz, no turbo boost
6gb OCZ Gold "low voltage" memory @ 1600mhz
Asus P6X58D Premium

Looks like you take the top spot for single-socket.

[botld92z]

Sweet. :-) I'll play more when I get home.

[StevensDE]

Nice! A new machine?

Yes, my old precision t7400 with 2x E5420 was broken-down and Dell changed it to a brand new precision t7400 with 2x X5450

[rge]

4.83ghz with 5C ambients/water cooled with new version, (6gb ram 7,7,7, 21, turbo off)

spi1M


spi32M


25K


50K

[botld92z]

^ nice!

[rge]

^ nice!

Thanks....only I have to stick to lower size benches, so the dual sockets dont completely hammer me even at their stock speed.

[poke349]

@RGE
Looks like I've got some updating to do on my website...

Since you clearly demolished the single-socket rankings for those sizes...

[poke349]

Updated the records on my site.
http://www.numberworld.org/y-cruncher/#FastestTimes


I just ran some benchmarks on my laptop...
And I have to say that the turbo boost is aggressive enough to "utterly destroy" the multi-core scalability of the program. (In a good way )

Under 4-core hyper-threaded intensive load, turbo won't activate at all.
+1 (1.73 GHz) is the max specification, but I only see it when it's under a less intensive load (by running less intensive stuff, or by disabling HT).

1.6 GHz under 8-thread load:



Under single-threaded load (locked to a single-core via processor affinity), it gets all the way to +8 turbo (2.66 GHz).
+9 (2.80 GHz) is the max specification, but again, I only see it when it's under a load that isn't as cpu-intensive... (such as compiling code)

2.66 GHz under 1-thread load:



And here are the results of single-thread vs. multi-threaded:


Only 2.64x faster with multi-threading. (whereas the desktop i7s will get more than 4x at this size)


Assuming turbo boost gets more and more aggressive in the future, we'll be seeing more of this.


*I was originally gonna make a specially optimized binary for this laptop (x64 SSE4.1 ~ Akari), but seeing as how temps will reach almost 90C under y-cruncher load, I don't think I'm gonna do it.
(Depending on the amount of ram, it usually takes about 12-48 hours of 100% cpu across all cores to tune for any particular machine... )

It's too nice of a laptop to destroy at such a early age. (I love this thing... Thanks 舅舅!!! Gonna get an SSD for it and bump the HD to the second HD slot. )

[Calmatory]

What kind of tool are you using for the development?

[poke349]

What kind of tool are you using for the development?

That's kind of a broad question. So I'm not sure what you're asking for.


Software:

I use Visual Studio and its compiler for the vast majority of the coding.
But all tuning is done using the Intel Compiler. The final SSE binaries are all compiled using the Intel Compiler while x86 binary is compiled using Visual Studio.
I have yet to learn how to use vtune and other profilers yet.


Hardware:

I have full access to about 8 or 9 different computers with which to run and test things on. (And a lot more if you include all my friends who let me toy with their machines.)
In addition to those, I also have access to a lot of the computers in the EECS department at my school here.

EDIT: But for the most part, there really is only one that matters. And that is obviously my 64GB workstation - which is the only thing I have that can run the larger tests.

[Kurumi]

Cool

btw, you need the rest of the stuff at the top of file to complete the batch validation. But I'll accept these anyway.

(Not like I've been validating any benchmarks anyway... )


EDIT:
What frequency is this at? (Before and after turbo boost)
And I'm assuming Core i7 920 and 12GB of ram?

Same specs as before. :P

[bonis62]

MajorGeek

[poke349]

Is this a mis-post? Since it doesn't seem relevant to this thread. Or any sort of digit-crunching for that matter...

[bonis62]

Is this a mis-post? Since it doesn't seem relevant to this thread. Or any sort of digit-crunching for that matter...

PrimeCores is written in assembler ( 85M in 11 secs at 3.2 GHz ) and does not use
advanced mathematical routine therefore fully
compatible with older processors,
the problem of using advanced mathematics is a bottleneck more than help, the good is writing custom advanced math routine that run on all processors,
for benchmarks the compatibility is crucial,
at least I think so.

[poke349]

PrimeCores is written in assembler ( 85M in 11 secs at 3.2 GHz ) and does not use
advanced mathematical routine therefore fully
compatible with older processors,
the problem of using advanced mathematics is a bottleneck more than help, the good is writing custom advanced math that run on all processors,
for benchmarks the compatibility is crucial,
at least I think so.

uh...

PrimeCores finds prime numbers. It has nothing to do with Pi.

y-cruncher is fully compatible with older processors. It just won't be as fast.

In my opinion, the true benchmark is one that adapts to the processor that it's running on by utilizing all special features it has. (Otherwise, why have those special features in the first place?)

Also, y-cruncher wasn't originally written for benchmarking.
It was done for a completely different purpose and later converted into a benchmark.

[bonis62]

uh...

PrimeCores finds prime numbers. It has nothing to do with Pi.

y-cruncher is fully compatible with older processors. It just won't be as fast.

In my opinion, the true benchmark is one that adapts to the processor that it's running on by utilizing all special features it has. (Otherwise, why have those special features in the first place?)

Also, y-cruncher wasn't originally written for benchmarking.
It was done for a completely different purpose and later converted into a benchmark.

if the code performs different operations for different CPU ,
you dont have a true benchmark, you have a simple test.

PI use floating point unit , Prime use integer unit,

I do not want put this into question,
do not speed race,
however, are convinced that a benchmark should run with the same (identical) code on many CPUs to be true benchmark.