I found a good collection regarding the ATI grayscreen problem (which spans from HD4 to HD5 series):
Granted, it is over 6 months old but current nonetheless.
It just hit a friend of mine with a brand new box with a brand new card.
Interesting read.
I’ve had all my machines with ATi cards (apart from a blue moon with some old nvidea geforce, which is now collecting dust somewhere in a drawer). Apart from the slowness episode we had a few years ago (glSelect on blendernation here http://www.blendernation.com/ati-slowdown-explained/ ) the cards have in my usage been very stable.
Currently on my main desktop I’m running two ATi HD 5550’s on software crossfire (through the motherboard ASUS M4A78T-E).
I’d link to the supported hardware page too, but as our wiki is currently down, I will just mention that a lot of Blender developers are using ATi cards - almost the majority IIRC.
/Nathan
The only thing that really sucks is the CUDA only software that seems to be out there. Whats the story with regards of OpenCL (as I also use smallluxgpu) between ATI and Nvidia ?
Small rectification - not the majority, just many 
/Nathan
OpenCL maintained by the Khronos group should work on both alike.
So does OpenGL, at 100% the stuff causing troubles are the OpenGL extensions, but the base versions work like charms on both manufacturers cards.
I think OpenCL is too young though to nail that.
What you can say though is the Power of the cards in FLOPS.
For those not familiar, a FLOPS = FLoating point Operations Per Second
So the brute calculation force, where you have to differe single and double precision.
Single GPU flagships:
HD5870: SP[2.720 GFLOPS] - DP [544 GFLOPS]
GTX480: SP[1.345 GFLOPS] - DP [168 GFLOPS]
That shows you that the HD5870 is superior to the GTX480 in openCL.
CUDA you´r screwed because the HD can´t play that game.
MS DirectCompute should be doable with both (DX11 IIRC)
That has a simple reason, nvidia plays ass again. The GF100 got a choke in the hardware to slow GPU computing, to sell their Tesla series, which is nothing else but the GF100 without hte choke, for the big buck to corporate customers, for that alone one should kick their marketing department in the kneecaps.
CUDA got its highlights - visit octane forum and post why not openCL :rolleyes: and you´ll be told.
I don´t like proppy stuff either, I like it open, but as enduser the choise sometimes isn´t yours.
So for OpenCL i´d get a Radeon
For CUDA a nvidia - dooohh.
For DirectComputing… well. I doubt anyone will use it excessively…
Doesn´t matter IMO if many or most. Actually it´s a good thing and even a few would be cool, as experience showed me that stuff working on ATI always works on nvidia, but sometimes not vice versa - in both ogl and dx.
Thanks for the info! :yes: I think I will holding of until a newer generation. My 5770 works fine except for the incompatibility CUDA stuff and some little problems once in a while, but that isn’t a big deal for the moment.
5770 has 1360 GFLOPS single precision, it can´t calculate double precision though.
So with OpenCL and single precision you should be in the GTX480 range.