[Bf-cycles] cycles, some questions about GPU

[Brecht Van Lommel]

Hi,

I guess this can almost turn into a GPU FAQ :)

2011/11/16 Sebastian Röthlisberger <lalli at gmx.net>:
> Cycles GPU Rendering currently is limited by the amount of available memory on the GPU, as far as i know.
> 1.-will cycles be able to render scenes bigger than the available Memory on (one) GPU?
>        1b.-if, in future, cycles uses multiple GPUs will it be still limited by the max amount of memory on one GPU or by the sum of all?

Rendering scenes bigger than available memory is planned to be added,
but it may be a year or so before I look at that. Multiple GPU's will
be limited by the max amount memory of 1 GPU.

> 2.-Is the rendering speed mainly determined by GFLOPS?
>        2b.-if not, can you tell me what is?

There is no single GPU feature that correlates directly to rendering
performance, only way to know is to test.

> 3.-Does Cycles make use of any "newer" NVIDIA technology? I mean will an older GPU (cycles ready) with same specs as a theoretical new One render at about the same speed?

It doesn't use any specific new functionality, but it benefits from
improved architecture of newer fermi cards.

> 4.-Cycles doesn't care about Nvidia SLI, "just" (will) use all GPUs, right?

SLI should probably be disabled to make multi GPU work even, not sure.

> 5.-Will cycles care about different GPUs (for example 1xOlder Cuda, 2xNew Cuda)
> 6.-Will/might it be possible to mix Cuda and OpenCL GPUs?

Probably using the same GPU's is more reliable, since you get the same
memory limits and GPU capabilities. If you've got an older GPU, use
that for drawing and the others for compute.

> 7.-Is there a chance that, in future, currently older GPUs like gtx 295 will drop out of Cycles support?

It's possible that some features will not be supported on them.

> 8-. just for private use: Do you recommend any GPU?

NVidia GTX 460, 560 and up, with as much memory as possible.

> Just wondering:
> 9.-I know you can't tell, but when you think will multiple GPUs be supported (CUDA)?

In the next few months.

> 10.-Do you know/predict anything about significant speed differences (of cycles) between CUDA and OpenCL?

There might be, we'll have to see how far we can get. In principle the
languages are similar and the hardware the same, but in practice it's
impossible to predict how it will turn out.

> 11.- Between 0-100%, approximate about where is cycles development at the moment.

There is no 100%, it's like asking when Blender will be done. Just try
it and see if it supports what you need.

Brecht.