LL with OpenCL - Page 21

LaurV · 2013-10-02, 13:32

Second clLucas test finished with success. I think this is a wonder, and either cllucas is more stable than cudalucas, or (and it hurts me to say it!

) amd cards are more stable than nvidia cards. I say that because the computer suffered ove 100 restarts and blue-screens in this period (not related to cl-stuff nor video card, but to other chain of things this computer is part now). Honestly I expected a mismatch, but I let it run...

I will dare now to start the third... (@kracker

)

nucleon · 2013-10-06, 12:22

I have 7990 (375W) 2xGPUs @1000MHz clock/1500MHz mem.

Using M30583963 as a test with FFT=2M I get iteration times around 3.9ms on both GPUs.

GPU0

Code:

Iteration 90000 M( 30583963 )C, 0x1bbfddfc1ddbe19f, n = 2097152, clLucas v1.01 err = 0.001099 (0:39 real, 3.8968 ms/iter, ETA 33:00:13)
Iteration 100000 M( 30583963 )C, 0xde1a5280cc5ab5cc, n = 2097152, clLucas v1.01 err = 0.001099 (0:39 real, 3.8989 ms/iter, ETA 33:00:39)
Iteration 110000 M( 30583963 )C, 0x5cfaabf2bb0e0487, n = 2097152, clLucas v1.01 err = 0.001099 (0:39 real, 3.8959 ms/iter, ETA 32:58:27)

GPU1

Code:

Iteration 90000 M( 30583963 )C, 0x1bbfddfc1ddbe19f, n = 2097152, clLucas v1.01 err = 0.0009766 (0:39 real, 3.8872 ms/iter, ETA 32:55:20)
Iteration 100000 M( 30583963 )C, 0xde1a5280cc5ab5cc, n = 2097152, clLucas v1.01 err = 0.0009766 (0:39 real, 3.8944 ms/iter, ETA 32:58:20)
Iteration 110000 M( 30583963 )C, 0x5cfaabf2bb0e0487, n = 2097152, clLucas v1.01 err = 0.0009766 (0:39 real, 3.8989 ms/iter, ETA 32:59:58)

Contrasting to my titan I get:

Code:

Iteration 470000 M( 30583963 )C, 0x64208421d1c227ee, n = 1835008, CUDALucas v2.03 err = 0.0283 (0:17 real, 1.7324 ms/iter, ETA 14:29:22)
Iteration 480000 M( 30583963 )C, 0x29d2686b8bb60915, n = 1835008, CUDALucas v2.03 err = 0.0283 (0:18 real, 1.7334 ms/iter, ETA 14:29:36)
Iteration 490000 M( 30583963 )C, 0x3308faaa69d0eef7, n = 1835008, CUDALucas v2.03 err = 0.0283 (0:17 real, 1.7254 ms/iter, ETA 14:25:18)

-- Craig

LaurV · 2013-10-06, 12:49

That is "normal". Your Titan uses a much shorter (therefore faster) FFT for this exponent. That shorter FFT is not optimized for cl-FFT. People are still working to convince clLucas to deal with non-powers-of-two FFT size (think when cudaLucas switched to v1.48 to 1.69 than later to 2.0, last year). Therefore, 1M9 FFT is much slower now, for clLucas, the same test would take 80 hours on your card. Therefore the 2M09 FFT is used, which is power of two. Try testing an exponent on your Titan which uses a comparable FFT size (ex: 38M exponent). Then, the comparison will be more accurate. Therefore posts like this.

(edit: by the way, my 580 is about 3% faster with 2097152, comparing to 1835008, you can try it for your Titan, an may get a speedup even for that 30M expo you are using for testing)

kracker · 2013-10-06, 16:40

Quote:

Originally Posted by nucleon

I have 7990 (375W) 2xGPUs @1000MHz clock/1500MHz mem.

Using M30583963 as a test with FFT=2M I get iteration times around 3.9ms on both GPUs.

GPU0

Code:

Iteration 90000 M( 30583963 )C, 0x1bbfddfc1ddbe19f, n = 2097152, clLucas v1.01 err = 0.001099 (0:39 real, 3.8968 ms/iter, ETA 33:00:13)
Iteration 100000 M( 30583963 )C, 0xde1a5280cc5ab5cc, n = 2097152, clLucas v1.01 err = 0.001099 (0:39 real, 3.8989 ms/iter, ETA 33:00:39)
Iteration 110000 M( 30583963 )C, 0x5cfaabf2bb0e0487, n = 2097152, clLucas v1.01 err = 0.001099 (0:39 real, 3.8959 ms/iter, ETA 32:58:27)

GPU1

Code:

Iteration 90000 M( 30583963 )C, 0x1bbfddfc1ddbe19f, n = 2097152, clLucas v1.01 err = 0.0009766 (0:39 real, 3.8872 ms/iter, ETA 32:55:20)
Iteration 100000 M( 30583963 )C, 0xde1a5280cc5ab5cc, n = 2097152, clLucas v1.01 err = 0.0009766 (0:39 real, 3.8944 ms/iter, ETA 32:58:20)
Iteration 110000 M( 30583963 )C, 0x5cfaabf2bb0e0487, n = 2097152, clLucas v1.01 err = 0.0009766 (0:39 real, 3.8989 ms/iter, ETA 32:59:58)

Contrasting to my titan I get:

Code:

Iteration 470000 M( 30583963 )C, 0x64208421d1c227ee, n = 1835008, CUDALucas v2.03 err = 0.0283 (0:17 real, 1.7324 ms/iter, ETA 14:29:22)
Iteration 480000 M( 30583963 )C, 0x29d2686b8bb60915, n = 1835008, CUDALucas v2.03 err = 0.0283 (0:18 real, 1.7334 ms/iter, ETA 14:29:36)
Iteration 490000 M( 30583963 )C, 0x3308faaa69d0eef7, n = 1835008, CUDALucas v2.03 err = 0.0283 (0:17 real, 1.7254 ms/iter, ETA 14:25:18)

-- Craig

Try using a greater -c and -aggressive(if not currently) and see if that changes anything.

@LaurV: I'll ask again, try pushing your memory clock up!

nucleon · 2013-10-07, 01:04

Aaah, the power of 2 FFT issue.

Ok, next time I get a chance to experiment I'll play with something more suitable.

-- Craig

Robish · 2013-10-08, 15:07

Quote:

Originally Posted by kracker

4 DC's finished here

Code:

M( 30766511 )C, 0x1ff14c8237b5e935, n = 2097152, clLucas v1.00
M( 30822937 )C, 0x1c656da41a256c21, n = 2097152, clLucas v1.01
M( 30888499 )C, 0xc296d9ac47d90339, n = 2097152, clLucas v1.01
M( 30976273 )C, 0x6c0367ea40d74647, n = 2097152, clLucas v1.01

My first CLLucas ;-)

M( 58191149 )C, 0x9108992abb23c5d1, n = 4194304, clLucas v1.01

10.5 days on 7870 with aggressive

More on the way....

kracker · 2013-10-08, 15:39

Quote:

Originally Posted by Robish

My first CLLucas ;-)

M( 58191149 )C, 0x9108992abb23c5d1, n = 4194304, clLucas v1.01

10.5 days on 7870 with aggressive

More on the way....

Nice.

I may put that on Prime95 for DCing, just for fun...

Robish · 2013-10-08, 16:53

Quote:

Originally Posted by kracker

Nice.

I may put that on Prime95 for DCing, just for fun...

Cool,

Next due in 14hrs, one in 24hrs and one in 36hrs.

Just bigger numbers take longer..... :-)

On a different subject, has many of you tried 100 million attempts?

Trying with Cudalucas at the mo.

1st attempt was reading 4500 hrs (190 days) on 20971520 (recommended by someone ;-( but after following this thread I tried 2097152 and (guess what) it dropped to 19 days on a gtx 690.

Should have results in two days time...

Just wondering if many were trying them yet?

Cheers

Rob.

Prime95 · 2013-10-08, 17:52

Quote:

Originally Posted by Robish

On a different subject, has many of you tried 100 million attempts?

1st attempt was reading 4500 hrs (190 days) on 20971520 (recommended by someone ;-( but after following this thread I tried 2097152 and (guess what) it dropped to 19 days on a gtx 690.

Should have results in two days time...

You cannot test a 100M digit number with an FFT length of 2 million. You may as well throw that result away. You will need an FFT length of at least 18 million.

kracker · 2013-10-08, 18:33

Quote:

Originally Posted by Prime95

You cannot test a 100M digit number with an FFT length of 2 million. You may as well throw that result away. You will need an FFT length of at least 18 million.

I'm suprised CULu allowed it?

Robish · 2013-10-08, 19:46

Quote:

Originally Posted by Prime95

You cannot test a 100M digit number with an FFT length of 2 million. You may as well throw that result away. You will need an FFT length of at least 18 million.

Really? Can you explain why? or a link to somewhere/thread that covers this please? It all looks like its going fine at the moment anyway......??

2013-10-06, 12:49	#223
LaurV Romulan Interpreter Jun 2011 Thailand 9653₁₀ Posts	That is "normal". Your Titan uses a much shorter (therefore faster) FFT for this exponent. That shorter FFT is not optimized for cl-FFT. People are still working to convince clLucas to deal with non-powers-of-two FFT size (think when cudaLucas switched to v1.48 to 1.69 than later to 2.0, last year). Therefore, 1M9 FFT is much slower now, for clLucas, the same test would take 80 hours on your card. Therefore the 2M09 FFT is used, which is power of two. Try testing an exponent on your Titan which uses a comparable FFT size (ex: 38M exponent). Then, the comparison will be more accurate. Therefore posts like this. (edit: by the way, my 580 is about 3% faster with 2097152, comparing to 1835008, you can try it for your Titan, an may get a speedup even for that 30M expo you are using for testing) Last fiddled with by LaurV on 2013-10-06 at 13:01

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfakto: an OpenCL program for Mersenne prefactoring	Bdot	GPU Computing	1676	2021-06-30 21:23
Can't get OpenCL to work on HD7950 Ubuntu 14.04.5 LTS	VictordeHolland	Linux	4	2018-04-11 13:44
OpenCL accellerated lattice siever	pstach	Factoring	1	2014-05-23 01:03
OpenCL for FPGAs	TObject	GPU Computing	2	2013-10-12 21:09
AMD's Graphics Core Next- a reason to accelerate towards OpenCL?	Belteshazzar	GPU Computing	19	2012-03-07 18:58

2013-10-02, 13:32	#221
LaurV Romulan Interpreter Jun 2011 Thailand 7²×197 Posts	Second clLucas test finished with success. I think this is a wonder, and either cllucas is more stable than cudalucas, or (and it hurts me to say it! ) amd cards are more stable than nvidia cards. I say that because the computer suffered ove 100 restarts and blue-screens in this period (not related to cl-stuff nor video card, but to other chain of things this computer is part now). Honestly I expected a mismatch, but I let it run... I will dare now to start the third... (@kracker )

2013-10-07, 01:04	#225
nucleon Mar 2003 Melbourne 5×103 Posts	Aaah, the power of 2 FFT issue. Ok, next time I get a chance to experiment I'll play with something more suitable. -- Craig