Re: NVidia Tesla K40

From: O. Hartmann <ohartman_at_zedat.fu-berlin.de>
Date: Fri, 31 Oct 2014 19:35:37 +0100
Am Fri, 31 Oct 2014 16:46:15 +0000 (UTC)
John Dison <jdison16_at_yahoo.com> schrieb:

> Hello!
> I want to use NVidia Tesla K40 GPU for parallel computing.Does FreeBSD support such a
> hardware? Thanks a lot!
> _______________________________________________
> freebsd-current_at_freebsd.org mailing list
> http://lists.freebsd.org/mailman/listinfo/freebsd-current
> To unsubscribe, send any mail to "freebsd-current-unsubscribe_at_freebsd.org"


You're out of luck. No, FreeBSD doesn't support GPGPU computing on modern graphics
hardware. Since 2009 we try to utilize FreeBSD for such a purpose, especially with
the use of OpenCL for high performance computing.

Around 2010 there was the glimpse of a dawn with Pathscale, offering a OpenACC-precursor
with their compiler infrastructure, supposed to be capable of utilizing FreeBSD and
nVidia GPUs (that time Tesla) with Pathscale-libraries (I forgot the name, it's CAPS or
something like that) but at the end it revealed its self as a heap of unfinished,
unmature code and by now the company or ist successor doen't even provide a commercial
product for FreeBSD - the project died!

It is even worse: FreeBSD ports dropped Nouveau! There are effords bringing libclc, the
new Mesa library, Glover and OpenCL in combination with LLVM 3.4/3.5 together to provide
OpenCL and via LLVM the nVidia PTX backend, but as far as I know this work is still under
development and immature and FreeBSD is no longer part of the scene due to the drop of
the Nouveau driver (I was told FreeBSD lacks important kernel features known for years
for KMS).

CUDA is completely out of view: nVidia doesn't provide support for FreeBSD. nVidia
fellows claim there is no need/request from FreeBSD people. I believe it is simply
ignorance and a stupid political issue, made years ago, that we suffer from by now.

The situation with open source AMD driver (radeonSI) is unknown to me. We made very bad
experiences with AMD hardware in the era of the AMD HD46XX/47XX and HD48XX chipsets and
the development is/was behind the recent chips available on market. On Linux there were
reports of successfully running OpenCL kernels on the radeonSI drivers with Glover/Mesa
and other mandatory software not available for FreeBSD - the FreeBSD X11 base system is
also "methusalem" and behind the recent development. If you would go with AMD, you would
have to stay with open source since AMD doens't provide BLOBs for their GPUs for FreeBSD.
For Linux, the AMD OpenCL SDK is very nice, their hardware is pretty fast (> nVidia!)
while their driver tend to crash. But as said, Linux only.

With nVidia you get a pretty fast and nice BLOB for all FreeBSDs and modern GPUs from
nVidia, but there is no OpenCL backend lib (or CUDA). Either way, it is a dead end with
FreeBSD.

There was also once a setup with the FreeBSD Linuxulator - but this is 32bit only and a
no go for serious scientific compuations. We also tried this with more modern FreeBSDs
(8.X, 9.X, 10-CURRENT that time), but with CUDA > 4 you're out of luck. And FreeBSD
doesn't have a 64bit support in the Linuxulator. And why running a Linuxulator?

If you really plan to use a beast like K40, consider dropping FreeBSD and using Linux! We
started with Ubuntu and OpenSuse that time. At the first moment the shock is heavy, but
you'll come along with Linux very soon. The only bad thing is the missing ZFS support as
someone might be used to in FreeBSD but this is also only a question of time (more short
than long!). Linux development is pretty fast these days, support is excellent and Linux
suffers from bad karma from the days of the Linux 2.4 kernel. That has changed
dramatically. A department developing security facilities is now dropping OpenBSD in
favour of Debian Linux, even OpenBSD is still considered more "bullet proof". But the
decision was made due to a dramatic driver issue - I mention this here because also
FreeBSD seems to suffer increasingly (compared to to Linux) from driver issues for high
performance equipment (special interlink adaptors, WiFi NICs and even the GPU issue!).

I'm very sad having no better experience to tell.

oh 

Received on Fri Oct 31 2014 - 17:35:52 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:53 UTC