GPU RAM.

Aramis Wyler · 2013-02-07, 14:47

I upgraded my video card recently to a gtx580, and was appalled to discover that the thing has 3 gigs of ram on it. Do any of the gpu program types (ll, p-1, tf, etc) take better/more advantage of on card memory than others?

firejuggler · 2013-02-07, 14:49

Unfortunatly, only P-1 might, and there is no software written for it (involving GPU) yet.

kracker · 2013-02-07, 16:26

LL does, if the range is high.

EDIT:CudaLucas of course

Aramis Wyler · 2013-02-07, 16:36

What constitutes 'high'? on p95 the memory needed never goes above 200 megs even when ll'ing ion the 60M block. Does CUDALucas use more memory on the card (not shared memory) than p95 does on the mainboard, or by high ranges are you talking the 332M range?

Dubslow · 2013-02-07, 20:31

Quote:

Originally Posted by Aramis Wyler

What constitutes 'high'? on p95 the memory needed never goes above 200 megs even when ll'ing ion the 60M block. Does CUDALucas use more memory on the card (not shared memory) than p95 does on the mainboard, or by high ranges are you talking the 332M range?

Both, I think. CUDALucas' memory footprint is larger than Prime95's footprint, but not by an order of magnitude. So you'd still only notice differences in the (e.g.) 332M range.

chalsall · 2013-02-07, 20:36

Quote:

Originally Posted by Dubslow

Both, I think. CUDALucas' memory footprint is larger than Prime95's footprint, but not by an order of magnitude. So you'd still only notice differences in the (e.g.) 332M range.

Just putting this out there...

There are many problem spaces which require a lot of near RAM.

Video games aside, having a GPU with a lot of RAM available is rarely a bad thing....

LaurV · 2013-02-08, 03:05

@OP: can you try running cudalucas -cufftbench with some very large FFT sizes and see where it stops? That is the limit for your exponents. For a "normal" GTX580 (the one with 1536MB) the "crash point" is somewhere at 12M FFT size, IIRC, and I was never able to run LMH on those (332M expo would need at least 18M FFT size).

Aramis Wyler · 2013-02-08, 13:51

Sure thing, I'll run it when I get home from work, in about 9 hours. Any particular increment? 1k?

LaurV · 2013-02-08, 14:27

Quote:

Originally Posted by Aramis Wyler

Sure thing, I'll run it when I get home from work, in about 9 hours. Any particular increment? 1k?

You will cry near the computer

Those large FFT need ages to go, I would suggest something more like a binary search, start with a 100k, or so, see where it crashes, do that range with 10k, etc. Or, depends on your time...

Aramis Wyler · 2013-02-08, 15:30

Ha! Fair enough. I'll run 11534336 20971520 1048576 just to see if it's even feasible to work in that range and if so work on something more in the 18-19M range at smaller increments (maybe 32k if I don't collapse into a river of tears).

I'll post the first search raw, and post the second one after I filter it down to remove the lower increments w/higher runtimes.

Redarm · 2013-02-08, 15:46

.

2013-02-07, 14:47	#1
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 110101000₂ Posts	GPU RAM. I upgraded my video card recently to a gtx580, and was appalled to discover that the thing has 3 gigs of ram on it. Do any of the gpu program types (ll, p-1, tf, etc) take better/more advantage of on card memory than others?

2013-02-07, 16:26	#3
kracker "Mr. Meeseeks" Jan 2012 California, USA 37×59 Posts	LL does, if the range is high. EDIT:CudaLucas of course Last fiddled with by kracker on 2013-02-07 at 16:31

2013-02-07, 16:36	#4
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 424₁₀ Posts	What constitutes 'high'? on p95 the memory needed never goes above 200 megs even when ll'ing ion the 60M block. Does CUDALucas use more memory on the card (not shared memory) than p95 does on the mainboard, or by high ranges are you talking the 332M range? Last fiddled with by Aramis Wyler on 2013-02-07 at 16:36

2013-02-08, 03:05	#7
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 41·251 Posts	@OP: can you try running cudalucas -cufftbench with some very large FFT sizes and see where it stops? That is the limit for your exponents. For a "normal" GTX580 (the one with 1536MB) the "crash point" is somewhere at 12M FFT size, IIRC, and I was never able to run LMH on those (332M expo would need at least 18M FFT size). Last fiddled with by LaurV on 2013-02-08 at 03:08

2013-02-08, 15:46	#11
Redarm Apr 2012 Berlin Germany 33₁₆ Posts	. Last fiddled with by Redarm on 2013-02-08 at 16:08

2013-02-07, 14:49	#2
firejuggler "Vincent" Apr 2010 Over the rainbow 2³×5×73 Posts	Unfortunatly, only P-1 might, and there is no software written for it (involving GPU) yet.

2013-02-08, 13:51	#8
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 2³×53 Posts	Sure thing, I'll run it when I get home from work, in about 9 hours. Any particular increment? 1k?

2013-02-08, 15:30	#10
Aramis Wyler "Bill Staffen" Jan 2013 Pittsburgh, PA, USA 2³×53 Posts	Ha! Fair enough. I'll run 11534336 20971520 1048576 just to see if it's even feasible to work in that range and if so work on something more in the 18-19M range at smaller increments (maybe 32k if I don't collapse into a river of tears). I'll post the first search raw, and post the second one after I filter it down to remove the lower increments w/higher runtimes.