-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
The GPGPU stuff that is making the computation much more accessible,
but at least in my field (games - uh make that real-time interactive
graphical simulations) if you give me more throughput near video ram,
I'll be sorely tempted to (surprise) use it to make more or better
pictures.
The interesting thing about the GPU is that it exposes a (fairly)
strict streaming computation model in which the user really only has
control over the computation kernel, and very little control over the
iteration construct. That makes using the high levels of parallelism
relatively easy and efficient. The challenge is in expressing non-
trivial algorithms in streaming ways. Aaron Lefohn's Glift suite
(http://graphics.idav.ucdavis.edu/graphics/publications/print_pub?
pub_id=837) is a nice wrapper around that material for more general
data structures on the GPU. The downside is that you had better have
a C++ compiler that does templates well.
I'm really hoping that we can find a way to get our users off the C/C+
+ bandwagon (and that includes the high-level shading languages as
well) and using something that can express the required computations
more naturally. There are some promissing-looking functional
approaches, but there's a huge barrier to adoption if it doesn't look
like C.
Paul
On 5-May-06, at 9:01 AM, David Leimbach wrote:
On 5/5/06, Paul Lalonde <[EMAIL PROTECTED]> wrote:
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
Aw, but I'd claim all that fancy 3-D graphics stuff is real
computation :-)
But yeah, GPU abuse for general purpose computation is just plain
scary. I thank my lucky stars that there is plenty of FLOPS to go
around in the Cell's SPUs.
Eh, nvidia's working on making the GPUs more accessible (via
compilers, kind of like Cell) for more general purpose computation.
The problem with FPGA, GPU, and "non-local" coprocessing cores is
usually the moving of data to them fast enough. Cell shouldn't have
this problem and with the new hypertransport stuff coming out, it
looks like one can easilly do NUMA like things inter-chassis too.
I don't know if this is cost effective, but streaming parallelism to
special coprocessors can be a big win in HPC.
Dave
Paul
On 5-May-06, at 8:46 AM, erik quanstrom wrote:
> if i were doing real computation, i wouldn't use a gpu i'd use a
> cpu. ;-)
>
> - erik
>
>> 8G/s? Nowhere near enough. Enough for text, but try doing real
>> computation using that GPU...
>> PS3 is running 25G/s bi-directional. Those bits move.
>>
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.1 (Darwin)
iD8DBQFEW3W0pJeHo/Fbu1wRAoOzAJ9C4d5WBnPm4hH1scoknQI1sFfuTgCgqC9c
Ft6mIE9ogrlaD9ltrNkMmjg=
=qWgd
-----END PGP SIGNATURE-----
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.1 (Darwin)
iD8DBQFEW3ulpJeHo/Fbu1wRAtbrAJ0RH5SpW4ZIx0W7BZIh3QXCRXt5MwCfYLVG
4DsnaEAu+s0hp/wAVsJZ5+U=
=R5Qr
-----END PGP SIGNATURE-----
