[cfarm-users] Upcoming board: Nvidia Jetson AGX Xavier

[Chuck Atkins]

Given what the platform and it's intended use is I think having the
proprietary nvidia drivers on there is the most appropriate.  Using the
nouveau driver on it might be "neat" but likely of limited practical use.

- Chuck

On Sat, Mar 6, 2021 at 7:28 AM Baptiste Jonglez via cfarm-users <
cfarm-users at lists.tetaneutral.net> wrote:

> On 05-03-21, Andy Polyakov via cfarm-users wrote:
> > > We have an opportunity to host a developer kit of the Nvidia Jetson AGX
> > > Xavier board in the farm.
> > >
> > > See:
> > >
> > >
> https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-agx-xavier/
> > > https://en.wikipedia.org/wiki/Tegra#Xavier
> > > https://en.wikipedia.org/wiki/Nvidia_Jetson
> > >
> > > It has a recent 8-cores ARM CPU,
> >
> > "Non-Cortex" might be a more accurate description. I mean most ARM
> > processors you encounter are Cortex cores, and there are many variants,
> > some are naturally more recent than others. This Tegra processor on the
> > other hand is not based on Cortex and in a sense is first of a kind. By
> > "first" I mean that even though it's described as part of the Denver
> > family, one has to recognize that previous family members used binary
> > translation to proprietary instruction set. This one doesn't, it's
> > actual ARM ISA implementation.
>
> Interesting, I hadn't realized it was such a specific design.  It makes it
> even more interesting for the farm.
>
> https://en.wikipedia.org/wiki/Project_Denver
>
> > > 32 GB RAM, and a big GPU.
> >
> > Don't get me wrong, 512 CUDA cores *is* impressive in the context of a
> > single-board computer, but one can't really refer to it as "big."
> > Thousands of cores is big, 512 is really entry level:-) In essence one
> > can divide the amount of CUDA cores by 32 in order to get a more
> > down-to-earth number. By 32, because that's how many lanes are processed
> > with single instruction. In other words one can also view it as a
> > processor with 32*32=1024-bit registers.
>
> Yes, you're right, but it's still rather impressive given the form factor
> and price.
>
> > Most common way to use Nvidia GPUs for general-purpose computing is
> > CUDA, with OpenCL being another option. Does nouveau support any of
> > that? Not that I know of. My understanding is that nouveau is about
> > rendering and acceleration, not about general-purpose computing. Of
> > course you can say that cfarm wants to foster nouveau development toward
> > OpenCL, but then it would probably be unsuitable as a shared farm
> > computer. Because developers are likely to need privileged access to
> > reload kernel drivers, not to mention a tendency for often reboots:-)
>
> Ok, I thought that nouveau could possibly support OpenCL, but I'm not
> familiar with this.
>
> Otherwise, we would need Nvidia proprietary drivers & CUDA, or just don't
> use the GPU at all.
>
> Thanks,
> Baptiste
> _______________________________________________
> cfarm-users mailing list
> cfarm-users at lists.tetaneutral.net
> https://lists.tetaneutral.net/listinfo/cfarm-users
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.tetaneutral.net/pipermail/cfarm-users/attachments/20210312/1306c8a3/attachment.html>