LL with OpenCL - Page 37

Lorenzo · 2016-01-20, 07:56

lalera, could you please post result for 332220523? (according with this thread http://www.mersenneforum.org/showthread.php?t=13185)

lalera · 2016-01-21, 09:43

Quote:

Originally Posted by Lorenzo

lalera, could you please post result for 332220523? (according with this thread http://www.mersenneforum.org/showthread.php?t=13185)

hi,
the ETA for M332220523 is around 3210 hours !
this is too long for me !

Lorenzo · 2016-01-21, 13:07

Quote:

Originally Posted by lalera

hi,
the ETA for M332220523 is around 3210 hours !
this is too long for me !

Yes. Sure ) Just for statistics )

------------
By the way. Who assigned this avatar to me?

I have never used it.

kladner · 2016-01-21, 15:27

Quote:

Originally Posted by Lorenzo

Yes. Sure ) Just for statistics )

------------
By the way. Who assigned this avatar to me?

I have never used it.

The forum gods work in mysterious ways. You can change it.

VBCurtis · 2016-01-21, 17:35

Quote:

Originally Posted by kladner

The forum gods work in mysterious ways. You can change it.

But you shouldn't!

LaurV · 2016-01-22, 04:59

Quote:

Originally Posted by VBCurtis

But you shouldn't!

Because is cute! (no other reason).

msft · 2016-02-01, 13:19

Hi,
Took out FFT kernel version.
This version only demonstration(for programmer).
Same speed,But fun.

airsquirrels · 2016-02-20, 23:18

I did a round of updated performance benchmarks, which I am also forwarding in detail to James.

The summary is the Fury X is now very usable for LL testing (I have several doing DCLL work), however cuFFT/CUDALucas still has better performance on the Titan cards.

Outside of the Fury X, the HD7990 is surprisingly formidable with its 2:1 DP, especially if you can actually keep it cool.

nlurski · 2016-04-03, 21:23

Would we be able to add the ability for it to auto-get work to do?

Madpoo · 2016-04-05, 21:47

Quote:

Originally Posted by nlurski

Would we be able to add the ability for it to auto-get work to do?

The primenet API is documented here Primenet API so it really just takes someone with time to try and implement that.

The source for Prime95 could probably be looked at to clear up anything along the way.

bverka86 · 2016-05-03, 01:41

MSFT, Just wanted to toss a note to you about the FFT length selection algorithm. It has a habit of picking very poor FFT Lengths in my experience. Running on 3 R9 270X's, it likes to pick FFT's like 1728K, 2240K, 3600K etc, all of which perform TERRIBLY. I assume this is true for all GCN cards, but at least on the 270X and the HD 7790 (my only other DP capable GPU), FFT Lengths divisible by 512K perform much better.

So for example, if 1536K wasn't good enough for M35861153, 2048K performs at least 3x better than 1728K. I would not mind digging into this algorithm, but I've been sick for quite some time and haven't dug in too deeply. Is the entire FFT length selection done in clLucas.cpp?

2016-02-20, 23:18	#404
airsquirrels "David" Jul 2015 Ohio 11×47 Posts	I did a round of updated performance benchmarks, which I am also forwarding in detail to James. The summary is the Fury X is now very usable for LL testing (I have several doing DCLL work), however cuFFT/CUDALucas still has better performance on the Titan cards. Outside of the Fury X, the HD7990 is surprisingly formidable with its 2:1 DP, especially if you can actually keep it cool. Attached Thumbnails

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfakto: an OpenCL program for Mersenne prefactoring	Bdot	GPU Computing	1724	2023-06-04 23:31
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

2016-01-20, 07:56	#397
Lorenzo Aug 2010 Republic of Belarus 2×89 Posts	lalera, could you please post result for 332220523? (according with this thread http://www.mersenneforum.org/showthread.php?t=13185)

2016-04-03, 21:23	#405
nlurski Apr 2016 1 Posts	Would we be able to add the ability for it to auto-get work to do?

2016-05-03, 01:41	#407
bverka86 May 2011 2³ Posts	MSFT, Just wanted to toss a note to you about the FFT length selection algorithm. It has a habit of picking very poor FFT Lengths in my experience. Running on 3 R9 270X's, it likes to pick FFT's like 1728K, 2240K, 3600K etc, all of which perform TERRIBLY. I assume this is true for all GCN cards, but at least on the 270X and the HD 7790 (my only other DP capable GPU), FFT Lengths divisible by 512K perform much better. So for example, if 1536K wasn't good enough for M35861153, 2048K performs at least 3x better than 1728K. I would not mind digging into this algorithm, but I've been sick for quite some time and haven't dug in too deeply. Is the entire FFT length selection done in clLucas.cpp?