double precision float in GPUs?

[leSLIe]

i want to start using CUDA for programming some applications that i want, and i was wondering, do the current GPUs have double precision float?

 

[risqu3]

i was going to write something witty with www.lmgtfy.com but i couldn't find any info on it

 

[necrophile]

Not quite sure, don't think so as i don't think there is much need for it on gpu's or at least there didn't need to be.

 

[evilsofa]

In the wikipedia article for CUDA, there is this bit of info in the Limitations section:

"For double precision there are no deviations from the IEEE 754 standard. In single precision, Denormals and signalling NaNs are not supported; only two IEEE rounding modes are supported (chop and round-to-nearest even), and those are specified on a per-instruction basis rather than in a control word (whether this is a limitation is arguable); and the precision of division/square root is slightly lower than single precision."

I have not the slightest idea what it means.

 

[Biskquik]

as I understand it, CUDA does support double precision, but the amount of stream processors available capable of double precision is different between the Geforce 8000,9000s and the GTX260+.

I actually googled and found this guide:
http://www.ddj.com/architect/207200659

part 7 seemed to discuss double-precision a bit more.

I'm holding off a bit on GPU programming until we see OpenCL tools available. I'm sticking to CPUs and OpenMP.

 

[Elledan]

The answer is in the CUDA documentation. RTFM.

To make it easy though, the only nVidia GPU which supports double-precision FP is the GT200-based series, meaning the GTX260 and up.

 

[necrophile]

You mean the 250 is not a 200 based gpu

 

[risqu3]

You mean the 250 is not a 200 based gpu

shhh some people here would kill you for saying that

 

[alg7_munif]

You mean the 250 is not a 200 based gpu

But the name is GT*2*0

 

[trinibwoy]

If you want to program in CUDA and you're asking these kinda questions on a video card forum I don't have much hope for you

Read the documentation and hit up the cuda forum at nvidia if you're serious.

 

[pxc]

Simple answer: The GT200 series chips provides one double precision FPU for each SM group of 8 SPs. The original GTX260 has 24 DP FPUs, the Core 216 model has 27 DP FPUs and the GTX280 has 30 DP FPUs.