FreeBSD, GPGPU and OpenCL/CUDA
cbergstrom at pathscale.com
Sat Aug 21 09:01:22 UTC 2010
Svein Skogen (Listmail account) wrote:
> On 18.08.2010 18:20, "C. Bergström" wrote:
>> Hi Oliver,
>>> The problem behind the subject is a little bit frustrating, so I do
>>> not know were to start.
>> Yeah it's a pretty big problem, but I can say others are looking at it
>> and taking small steps in the right direction.
>>> First, and this hasn't changed since the last 15 years, FreeBSD lack
>>> in support of professional Compiler vendors. Pprtland Group offers
>>> only Linux compilers, as far as I know Intel does not offer a native
>>> FreeBSD 64 Bit compiler. So we are stuck with gcc and gfortran
> If we see beyond the CUDA part of this question, it should be noted that
> ATI/AMD has kept to their promise of actually supporting opensource.
> (see also
> I must admit not having tested that code myself (my two 5970s sit in a
> windows box), but projects should maybe consider the ATI/AMD cards.
> 928GFLOPS double-precision per card (4.64TFLOPS single precision) with
> proper documentation should at least give a proper start to things...
I'm sorry, but this is only a half truth.. The firmware they don't
release, despite what they want you to believe, is both interesting and
should imho be open. Also the documentation they've released clearly
has parts chopped out and when you actually try to do something useful
it's missing important details. (My perspective is biased since by
something useful I mean write an assembler or compute backend) I've had
one awesome engineer at AMD trying to help us fill in some of those
blanks, but if or when the public docs will be updated. Overall due to
ATI hw missing ECC ram, demand paging and a few other hardware features
I don't consider it a serious contender in HPC yet. Lastly, you'll see
the real efficiency sucks and you can't get anywhere near the peak
GFLOPS off those cards.
Oh.. and most important to the land of FBSD is that the ATI drivers are
still using the rather nasty TTM.. This is *not* very easily ported to
FBSD, blocks compute capabilities and our early benchmarks show kicking
it out has given nice performance increase in certain areas of graphics.
(Hope I don't come across negative... I'm just trying to give real
feedback based on our experience)
More information about the freebsd-questions