--- Timothy Miller <[EMAIL PROTECTED]> wrote:
> I'd like to give this a shot, but I see a chicken and egg problem. 
> To
> get them to take us seriously, we probably have to do significant
> work
> in their field.  But if we do that, there's no guarantee they'll
> offer
> us any funding.  How could we get them to offer us funding up front?

The weaknesses of GPUs are:

There is no support for integers or fixed point.  There is only
fixed-function graph reduction.  Output must be a two-dimensional
array.  There is no hardware support for doubles, although there are
SIGGRAPH presentations this year about emulating them.

If those can be beaten in a product costing less than the Cell, there
will be markets.  Scientific computing wants cheaper supercomputers,
financial and statistical computing wants faster APLs,[1] game
programmers[2] want to run "essentially functional" programs which
transform "a small input data set to a small output data set, making
use of large constant data structures."

AGEIA is testing the market for a product that solves a sliver of that
last problem.

> From: Jon Smirl <[EMAIL PROTECTED]>
> I don't see much mention of the concept of General Purpose GPU usage.
> Like what is happening on this website. http://www.gpgpu.org/ These
> people have a lot of money, millions of dollars to throw at hardware.
> They only want one thing and that is performance.

They also want accuracy.  The OGP has an advantage here in that ATi and
NVidia have historically come down on the wrong side of this issue in
their drivers and compilers.  The compilers can be fixed, but the idea
that your computation may do the wrong thing if your executable is
named quake.exe may be tougher to defeat.

> Have you consider doing a stripped down design. No DMA, no PROM, no
> specific 2D. Only DVI scanout, PCI, RAM and a shader CPU. Then put
> all
> of the effort into the shader CPU? When the shader CPU works, start
> copying it and see how many you can get on a chip.

The GPGPU architecture has three distinct processor types.  The first
is the graph reducer, usually the host CPU, which runs the overall
program and directs the other processors.  The second is the vertex
shader, which performs the array reshaping operations.  The Cell and
PhysX chips use DMA for this.  The third is the pixel shader, which
does the streaming computation and relies on the other two to give it
meaningful streams.

It is this architecture which gives the GPU its power, not the speed of
the particular shaders.  It isn't as easy as designing a good pixel
shader; there must be a good architecture to keep it busy.

As for how many shaders would fit, it's only a question of how much
money one has.  Everything the fragment processor needs can be supplied
by standard cells, and the fixed function graph reducer doesn't have a
scale limit.


[1] http://www-306.ibm.com/software/awdtools/apl/

[2] Slides:
http://www.st.cs.uni-sb.de/edu/seminare/2005/advanced-fp/docs/sweeny.pd
See slides 12 and 17.
Discussion: http://lambda-the-ultimate.org/node/1277


__________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around 
http://mail.yahoo.com 
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)

Reply via email to