Re: NVidia Tesla K40

Dear Jean-Sebastien,

just to chime in on one aspect I care about:

On Sun, Nov 2, 2014 at 11:25 AM, Jean-Sébastien Pédron <dumbbell_at_freebsd.org
> wrote:

> On 31.10.2014 17:46, John Dison wrote:
> > Hello!
>
> Hi!
>
> > I want to use NVidia Tesla K40 GPU for parallel computing.Does
> > FreeBSD support such a hardware?
>
> As O. Hartmann explained, we don't support GPGPU on NVIDIA hardware,
> neither with the binary driver because NVIDIA doesn't provide any
> libOpenCL.so for FreeBSD, nor with the FOSS driver because we don't have
> the kernel video driver.
>
> On the AMD front, GPGPU works with the open-source stack (libclc and
> Clover, Mesa's libOpenCL.so), but the Clover port isn't committed yet.
> We didn't have the time to polish it so far (only libclc is available in
> devel/libclc).
>
> Note that Clover is still a WIP. Here's a matrix of implemented and
> missing features:
> http://dri.freedesktop.org/wiki/GalliumCompute/
>
> When I tried it with simple programs found on the Net, it worked.
> However, I couldn't use it with graphics/darktable:
>     1) Currently, the amount of memory reported by Clover is hard-coded
>     2) darktable expects features not implemented
>
IMHO, a very good and relevant test is clBLAS (
https://github.com/clMathLibraries/clBLAS ). Anything HPC will require
these functions to work properly and (albeit computationally demanding)
they are well controllable for benchmarks (in terms of setup and memory
dimensioning).

Coming from a HPC background, the next step would be something like MAGMA (
http://icl.cs.utk.edu/magma/software/index.html ) which uses clBLAS and
adds some LAPACK-style functionality to it. Again, well controllable and
computationally challenging.

I am happy to act as tester, btw.

Looking forward to this landing in BSD, thanks so much for your and all of
the graphics teams work!

Johannes