CUDA and video memory

dbaugh · 2012-03-19, 21:55

What are the advantages of having a 3GB GTX580 versus a 1.5GB GTX580 for CUDA programming purposes? Any difference when running mfaktc?

firejuggler · 2012-03-19, 22:03

AFAIK, no advantage for the program used here . Shader clock is what matter (for mfaktc at least)

Dubslow · 2012-03-19, 22:21

If you look at GPU-Z while running mfaktc, you'll see just about zero usage (except for the screen, obviously). It's been said that CUDALucas uses a lot more memory, but more memory shouldn't make a difference, since the memory requirements are set by the FFT. If you were interested in creating a P-1 program for GPU, then yes, memory does matter for Stage 2. JUST for CUDA programming (i.e. application independent), I don't see why it would make a difference.

Karl M Johnson · 2012-03-20, 03:06

A lot of memory is very useful if you want to draw high-res fractals using your GPU.
But for mfaktc, it doesnt matter.

dbaugh · 2012-03-20, 03:44

Very helpful. Thank you.

KyleAskine · 2012-03-20, 10:45

It might matter if we ever get P-1 for CUDA.

henryzz · 2012-03-22, 22:58

I thick I might have seen that sometimes cards with more memory have more memory bandwidth.

LaurV · 2012-03-23, 05:55

Quote:

Originally Posted by henryzz

I thick I might have seen that sometimes cards with more memory have more memory bandwidth.

That is generally viceversa. If the chip has a wider bus, then more "parallel" memory chips are added, making a larger total of memory, but one can not judge the bus width (and the bandwidth) from the solely amount of memory on the card. For example, Fermi cards have from 1GB to 6GB of memory, but all have 384 bit bus, that is, they can transfer 48 bytes in the same time. Generally, the bandwidth is proportional with the bus width and the clock. Faster memory clock = larger bandwidth. Wider bus = larger bandwidth.

jasong · 2012-03-25, 05:48

Threads like this indicate why it's frickin annoying that Windows programs require a graphical interface to exist even if no one is at the computer. Or at least that's what I've heard.

KyleAskine · 2012-03-25, 12:51

Quote:

Originally Posted by jasong

Threads like this indicate why it's frickin annoying that Windows programs require a graphical interface to exist even if no one is at the computer. Or at least that's what I've heard.

Huh?

jasong · 2012-03-25, 14:15

Quote:

Originally Posted by KyleAskine

Huh?

Since graphics cards are used to maintain what's displayed on the monitor, more power is available if you can totally avoid the needs to run those particular programs that display stuff on the monitor. Unfortunately, some Windows programs won't run at all without running the part that displays the data. This is true even when there's no window open.

I don't believe it's possible to run Windows without at least one or two of the running programs running a graphical interface. Basically, the desktop you see exists even if no one is using the computer, even though it doesn't benefit the computer to maintain the desktop when no one is viewing it.

This is another reason why Windows is so damn inefficient for running data heavy programs. Sometimes, simply switching to Linux gives a huge improvement simply because it isn't Windows.

2012-03-19, 21:55	#1
dbaugh Aug 2005 121₁₀ Posts	CUDA and video memory What are the advantages of having a 3GB GTX580 versus a 1.5GB GTX580 for CUDA programming purposes? Any difference when running mfaktc?

2012-03-20, 10:45	#6
KyleAskine Oct 2011 Maryland 100100010₂ Posts	It might matter if we ever get P-1 for CUDA. Last fiddled with by KyleAskine on 2012-03-20 at 10:45

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Too many video cards	cardogab7341	GPU Computing	8	2014-12-26 03:32
Nvidia Cuda 6.0 unified memory management	Manpowre	GPU Computing	4	2013-11-28 18:53
An awe inspiring video...	chalsall	Lounge	0	2011-07-31 18:47
Video Poker	Unregistered	Information & Answers	7	2010-06-23 02:42
Use video accelerators for LL testing?	BOZCOM	Hardware	2	2004-07-30 15:45

2012-03-19, 22:03	#2
firejuggler "Vincent" Apr 2010 Over the rainbow 5435₈ Posts	AFAIK, no advantage for the program used here . Shader clock is what matter (for mfaktc at least)

2012-03-19, 22:21	#3
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3×29×83 Posts	If you look at GPU-Z while running mfaktc, you'll see just about zero usage (except for the screen, obviously). It's been said that CUDALucas uses a lot more memory, but more memory shouldn't make a difference, since the memory requirements are set by the FFT. If you were interested in creating a P-1 program for GPU, then yes, memory does matter for Stage 2. JUST for CUDA programming (i.e. application independent), I don't see why it would make a difference.

2012-03-20, 03:06	#4
Karl M Johnson Mar 2010 3×137 Posts	A lot of memory is very useful if you want to draw high-res fractals using your GPU. But for mfaktc, it doesnt matter.

2012-03-20, 03:44	#5
dbaugh Aug 2005 11² Posts	Very helpful. Thank you.

2012-03-22, 22:58	#7
henryzz Just call me Henry "David" Sep 2007 Liverpool (GMT/BST) 1011101100110₂ Posts	I thick I might have seen that sometimes cards with more memory have more memory bandwidth.

2012-03-25, 05:48	#9
jasong "Jason Goatcher" Mar 2005 110110110011₂ Posts	Threads like this indicate why it's frickin annoying that Windows programs require a graphical interface to exist even if no one is at the computer. Or at least that's what I've heard.