Re: Southport + Seeding FPGA application acceleration in our community
Heyho,
On 10/26/2011 05:18 PM, Johan Henriksson wrote:
> This is some nice development!
:) development this means, indeed!
> That said, there are some other companies that has tried this out and I
> doubt this will really take off.
Well, I do not see any FPGA company that has collaborated with the
Open Source community so much, yet ... well, indeed, there is one
in Sweden behind OpenRisc [1,2], but application acceleration is yet
not so much theirs from what I have understood. Someone please
provide the link, I could not find it on the pages any more.
> There might be some potential for the
> protein folding stuff where they still struggle, but the GPUs have taken
> whatever little market there was. They're already in the computer and as
> far as I have seen have more power/$
This depends all on the application. GPUs have an edge on floating
point operations. I can tell this with more confidence after our tutorial.
from what I understood, when you know a range yours floats to
be living in, you can come up with some more esoteric (but simpler)
representation of your floats which then beat the hell out of the GPUs
again.
It is however not completely unreasonable to have mixed GPU / FPGA
setups.
Somewhat funny that you mentioned "power" and you did that correctly,
indeed: FPGAs are just _far_ more energy efficient than GPUs.
> (so we really have to make sure the
> opencl/cuda drivers work!).
(I went through that pain recently for my ATI ones ... only to then learn
that my card does not feature doubles :o/ )
> When it comes to things like alignments and
> other routine operations, we live in an age when people take their perl
> code and run it on their 100 core cluster... A total waste of resources but
> biologists are not computer scientists and do not care about doing things
> The Right Way (tm). I would say that getting them to even use hashmaps and
> other basic techniques is usually out of reach.
This again is why we are needed as a distribution. When the preparation
of the FPGA for the computation is noticed then it will not be used.
We need to prepare for that all and think about how to extend the
regular packages either directly with upstream or in some other way
for an automated (aka painless) acceleration.
> but we live to see what happens :)
Yes, this is what it is - an experiment.
Glad to be a part of it
Steffen
[1] http://openrisc.net/
[2] http://opencores.org
Reply to: