What's CUDA ?
CUDA is a parallel processing language invented by NVIDIA that runs on the current (and most older) GeForce video card 'Graphic Processing Units' (GPU's). Due to the highly parallel nature of FFT's, SETI is an ideal application for running on a GPU. Even older card GPU's typically process SETI wu's "5 to 10 times faster" than any CPU.
ATI offer something similar for their cards, however (as of end 2011) SETI does not run on them
How do I get SETI to run on my Graphics card ?
The standard SETI Client will automatically look for, and run on, the GPU (instead of the CPU), if it can.
The SETI Client will also state the 'Processing Power' (GFlops) of your card
What version of CUDA do I need ?
SETI will run on CUDA v1.1 and above, i.e. just about any NVidia GeForce card.
The SETI Client code is 'Open Source', so 'optimised' SETI Clients exist that will take advantage of whatever capabilities your CPU / GPU have.
Can I run SETI on both the CPU and GPU ?
Currently, no
When running on the GPU, the CPU makes very little difference .. a low power single threading CPU with a decent GeForce CUDA card will outperform even the fastest quad-core HT DDR3 equipped CPU every time. So, if you have a PCIe-16 slot, don't waste money upgrading your motherboard (or CPU), just buy a GeForce instead !
How much Graphics card memory do I need ?
Generally, there is little cost difference between a 512 Mb and a 1Gb card - and little performance difference either. However, whilst currently the SETI client needs only 350 Mb or so of free graphics RAM (so will run, just, on a 384Mb 9800GSO), SETI used to run just fine on a 256 Mb card !
I thus advise avoiding both the 384Mb 9600GSO (get the 768Mb version) and all 512Mb cards (get the 1Gb version). There seems little doubt that the SETI Client will soon be 'improved' to the point where it won't run in 384Mb, and, sooner or later, even 512Mb won't be enough
Is it worth 'overclocking' the card ?
Yes & no. SETI performance follows the core clock (as opposed to the RAM / Bus clocks), so if you can boost the core clock 20% you will get 20% SETI improvement - however overclocked cards will require more power and a LOT more cooling - and a single glitch can crash the card and loose you whatever wu's you were processing.
Some cards overclock better than others - the GTX 260, for example, can go from about 500 GFlops to 650 GFlops or so by overclocking from 576/1242/1998 to 725/1100/1592 ... whereas the 9800GT can only be overclocked from 364 to 384 GFlops (whilst the 9800GTX/+ will do 470 GFlops without overclocking).
What's the 'best' graphics card ?
Depends ... are you looking for "GFlops per £" (wu's processed) or "GFlops per Watt" (heat) ?
GFlops per £ = your purchase price, whilst GFlops per Watt (heat) = your running costs
Generally, the 'older' the graphics card, the cheaper it is ... however the older it is, the less processing power per watt of heat you get. You should, of course, avoid any card that has just been released, since it will be sold at 'premium prices' to the 'extreme gamers' aka 'speed-freak morons' (or just plain 'mugs').
Note. Avoid the 'M' (mobile) versions ... they typically have half the processing power (& generate less than half the heat) but are sold at a much higher price than the desktop version.
Given the capabilities of the current 'top end' cards, which exceed 1 TFlops, you should be aiming at something at least half that performance (i.e. in the 500 GFlops region) for 1/4 the price. So 9800 GTX+ / GTS250** (the 'new name' for the 9800 GTX+) @ 470GFlops should be a good price/performance point to aim at (as of end 2011), as would the GTX260 @ 510 GFlops (if it can be had at a similar price).
**The GTS250 has a lower power consumption than the GTX+, so is a better choice, especially if you can get it at the same price as the GTX+ !
I recommend creating your own 'table' of "GFlops per £" - and then choose the best 'value' (or whatever you can afford)
Be very aware of each cards power requirements since you may have to add the cost of a new PSU to the "GFlop's per £" figure when comparing different cards.
Click 'Next >>' in the Navigation bar (left) for my Raspberry Pi projects