CUDA and using the computer

TimSorbet · 2010-06-12, 11:35

What is the impact of running a CUDA program on CPU usage and computer responsiveness? E.g. how much of a quad would be needed to keep a GPU busy? And, how usable is the computer while this is happening? Is it barely usable due to unresponsiveness (I think I remember hearing that's the case) or is it as responsive as a program running low priority CPU work?
Do any of the current CUDA prime/factoring programs exist for, or could be easily built for, 32-bit Windows? Or could they run through Cygwin without a significant speed decrease?

ET_ · 2010-06-12, 11:48

I used mfaktc to trial-factor huge exponents. The program runs on both Linux and Windows, and is 32 and 64 bits aware.

Of course, 64bit version is about 33% faster than 3bit.

Usually it charges only one core of a multicore CPU. I used mfaktc AND multi-threaded mprime on my Linux box without losing speed.

As for responsiveness, Windows (64) version is about 10% slower than Linux 64, but you can use your PC while CUDA operations are used. On Linux you can't, unless you run terminal version without Xserver.

Luigi

em99010pepe · 2010-06-12, 12:27

A Nvidia GTX 480 uses something like 500 W in load. A core i5 750 uses 200 W under load ( 4 cores). How much faster is the mfaktc CUDA version compared to the CPU version? I just want to know what's the best in terms of energy efficiency.

Carlos

TheJudger · 2010-06-12, 12:45

Hi Carlos,

GTX 480 + i750 + 3 instances of mfaktc: ~360W for the whole system.
The throughput depends on the exponent and factoring ranges. But I think a 8-10 fold throughput increase can be achived (compared to 4 cores running Prime95).
I took 3 100 million digits exponents (M332.235.xxx) from 2^70 to 2^77 in ~51 hours. 3 instances of mfaktc on GTX 480 + i7 750.

So running mfaktc on a GTX 465/470/480 is very energy efficient. Perhaps a GF104 based GPU is a even better deal but we have to wait until it is released.

Oliver

2010-06-12, 11:35	#1
TimSorbet Account Deleted "Tim Sorbera" Aug 2006 San Antonio, TX USA 10B7₁₆ Posts	CUDA and using the computer What is the impact of running a CUDA program on CPU usage and computer responsiveness? E.g. how much of a quad would be needed to keep a GPU busy? And, how usable is the computer while this is happening? Is it barely usable due to unresponsiveness (I think I remember hearing that's the case) or is it as responsive as a program running low priority CPU work? Do any of the current CUDA prime/factoring programs exist for, or could be easily built for, 32-bit Windows? Or could they run through Cygwin without a significant speed decrease? Last fiddled with by TimSorbet on 2010-06-12 at 11:35

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
CUDA 5.5	ET_	GPU Computing	2	2013-06-13 15:50
AVX CPU LL vs CUDA LL	nucleon	GPU Computing	11	2012-01-04 17:52
Best CUDA GPU for the $$	Christenson	GPU Computing	24	2011-05-01 00:06
CUDA P-1?	nucleon	GPU Computing	2	2010-11-17 17:52
CUDA?	Xentar	Conjectures 'R Us	6	2010-03-31 07:43

2010-06-12, 11:48	#2
ET_ Banned "Luigi" Aug 2002 Team Italia 11401₈ Posts	I used mfaktc to trial-factor huge exponents. The program runs on both Linux and Windows, and is 32 and 64 bits aware. Of course, 64bit version is about 33% faster than 3bit. Usually it charges only one core of a multicore CPU. I used mfaktc AND multi-threaded mprime on my Linux box without losing speed. As for responsiveness, Windows (64) version is about 10% slower than Linux 64, but you can use your PC while CUDA operations are used. On Linux you can't, unless you run terminal version without Xserver. Luigi

2010-06-12, 12:27	#3
em99010pepe Sep 2004 2·5·283 Posts	A Nvidia GTX 480 uses something like 500 W in load. A core i5 750 uses 200 W under load ( 4 cores). How much faster is the mfaktc CUDA version compared to the CPU version? I just want to know what's the best in terms of energy efficiency. Carlos

2010-06-12, 12:45	#4
TheJudger "Oliver" Mar 2005 Germany 5×223 Posts	Hi Carlos, GTX 480 + i750 + 3 instances of mfaktc: ~360W for the whole system. The throughput depends on the exponent and factoring ranges. But I think a 8-10 fold throughput increase can be achived (compared to 4 cores running Prime95). I took 3 100 million digits exponents (M332.235.xxx) from 2^70 to 2^77 in ~51 hours. 3 instances of mfaktc on GTX 480 + i7 750. So running mfaktc on a GTX 465/470/480 is very energy efficient. Perhaps a GF104 based GPU is a even better deal but we have to wait until it is released. Oliver