Radeon as CUDA device?

[Corbin Simpson]

On 07/10/2009 12:37 AM, Ancoron Luciferis wrote:
> OpenCL would be really great!
> 
> I know it's a big fat load of work but there needs to be something soon,
> because nVidia is much bigger already in the computing world. And using
> the Stream SDK is not of much use for many developers because they know
> that they have to do the work twice as nVidia's CUDA can't just be
> ignored - for economical reasons of the business.

Meh. "Needs" is such a weird word around here, because as far as I know
none of us are getting paid to do any GPGPU work.

> Doesn't AMD already work on OpenCL for their cards? Well if they do they
> should make it open source from the beginning, because otherwise the
> OpenCL implementation will have similar problems as their proprietary
> drivers have: some things just don't work in some situations.

OpenCL is non-trivial to support. It requires a lot of infrastructure,
and a lot of per-driver code since r300 and r600 are both kind of
experimental/non-working in Gallium right now.

> Roland Scheidegger wrote:
>> On 10.07.2009 00:06, wild-thing wrote:
>>> Hello there!
>>>
>>> I guess this could be a feature request: (An unusual question)
>>>
>>> I am sure you know BOINC: http://boinc.berkeley.edu/ I am spending
>>> processor time for one of the related projects as user only. And
>>> there is possibility to use a graphicsboard as a so called CUDA
>>> device with the BOINC client software.
>>>
>>> This is using the Graphicsboard as an additional hardware for faster
>>> calculations. Currently it seems to only work with NVidia boards? So
>>> here is my question:
>>>
>>> Are there plans of the xorg developement team to add possibility to
>>> use lets say my: "X1950 XTX" as a CUDA device with "radeon" driver?
>>>
>>> Can you implement this or would u at least, please?
>> CUDA probably doesn't map too well to radeon hardware. Plus you've got a
>> pre-r600 card which isn't really suited to that anyway. I guess there
>> would be interest in implementing OpenCL, but only for r600 and beyond,
>> but I don't think anyone is working on that already...
>> (Note that for closed source AMD drivers you can get the stream sdk
>> today, but this is restricted to cards newer than r600 too.)

We could totally do CUDA and CTM and all kinds of GPGPU on r300-r500
cards. It would just suck a lot.

Seriously, though, I expect that we'll get "free" r300 support in
Gallium for a few GPGPU libs, when those libs get regular Gallium
support. Just don't be surprised when it's all slow and fail.

~ C.