Can anyone recommend a card that has compute capability 2.0, and would work in a sysem with a 250 watt power supply? (And if possible in parallel with the current 8600 GT card?) All the ones I see seem to require a minimum 350 watts, and often more.
I do s/w development, so the CUDA code seldom has to run for more than a few minutes on my system. (Though in production it might run hours or more.) I need to be able to test versions at higher compute levels, and also to verify that the code runs on (and makes good use of) multiple GPUs.
I really don’t want to go to a new machine, because of the time I’d lose getting everything set up to my satisfaction. I also need something that is quiet, with no fan noise &c when I’m not actually running my code on the GPU.
(I apologize if this is the wrong place to ask, or if the question’s been answered already. For some reason, the search facility doesn’ let me search on words line “watts”.)
Do you want 2.0 specifically or will you be satisfied with 2.1?
The problem with 2.0 is that 2.0 comes only in high-end cards, and even the most basic of those can draw 200 watt from the power supply under full load. And that’s in addition to all the other hardware you have installed (CPUs often draw 50 to 100 watt, each hard drive adds 10 watt, your 8600 GT can draw 40-50 watt, etc.)
If you’re ok with 2.1, depending on the power draw from the rest of your system, you might be able to sustain GT430 or GT440. Especially if you underclock them.
2.1 would be fine, as it should be downwards compatible with 2.0. It’s just that some algorithms might work better/faster with some 2.x capabilities, but it’s hard to say without something to test on - and likewise with code using 2 or more GPUs.
After a bunch more searching, I think I may have answered my own question. Looks like a Quadro 600 card will draw about 60 watts or so. For some reason that didn’t show up in my earlier searches.
I’m curious about underclocking. I see quite a few boards advertised with overclocking, so thought that was something that had to be done in hardware?
Actually, the best thing for you to do is just to buy a new low-noise 750 watt power supply. The power supply can be swapped in 15 minutes with a screwdriver and you won’t need to reinstall any software.
Advertising boards with overclocking is mostly a marketing ploy. All they do is set the default values in BIOS and probably run some quick tests to make sure that the card is stable. But clocks remain fully adjustable in software in real time.
Quadro 600 is essentially a GeForce GT 430 with a “Quadro” label on it.
Waste heat should be the percentage of actual load, not peak load.
I have a fairly power-hungry PC (170 W video card, 95 W CPU, a lot of fast memory, a bunch of fans), powered by a middling 85%-efficiency 750 W power supply. The whole setup eats 80 W when idle, as measured at the outlet.
Sure - but I’m used to doing most of my work (which is mostly non-CUDA) on a notebook that (exclusive of the monitor) draws under 20 watts when I’m just doing my most-of-the-time text editing the source code. I suppose it’s a personality thing: waste & excess bother me, which is part of the reason I manage to get paid for making other people’s code run faster :-)
I bet you could fit a nVidia GTS 450 card into this PC without blowing the power supply. Use it for CUDA only, not for 3D games.
I blew an underpowered PSU once playing Cysis. Running CUDA was no problem with the same PSU, before it blew External Image
Here is the idle and graphics load power consumption of a complete PC for several graphics cards, and the GTS 240 system draws around 249 Watts under gaming load.
Seems reasonable: there should be significant chunks of silicon, like rasterizers, that aren’t used at all when running CUDA. And I don’t do gaming at all - sitting in front of the computer is !@# work :-(
