GPU (was: Putting squeak in business.)

Joshua 'Schwa' Gargus schwa at cc.gatech.edu
Tue Nov 18 01:19:37 UTC 2003


On Mon, Nov 17, 2003 at 09:28:01PM -0200, Jecel Assumpcao Jr wrote:
> On Monday 17 November 2003 20:25, Joshua 'Schwa' Gargus wrote:
> > One of my problems is that I'm a graphics geek... even if we can
> > somehow manage to replace the CPU + software VM with an FPGA VM
> > implementation, how would we replace the GPU?
> 
> It certainly not easy, just like doing a CPU in FPGA can be complicated. 
> Look at how awful a Z80 turns out to be when done that way.
> 
> At the very least you have to start out with a large FPGA with nice 
> features. The Spartan III 400, for example, has 16 hardwired 18x18 
> multipliers.
> 
> > Or could we still interact with commodity GPUs? 
> 
> That is an option, though any chip that requires a huge fan and extra 
> power lines (beyond those in the AGP connector) isn't something I will 
> be seriously considering.

Agreed.  However, it's usually only the top-end cards that require this.

> 
> > I recall a message from you some time
> > ago that revealed that you have put a significant amount of thought
> > into graphics on your type of system.  Have your thoughts changed at
> > all now that we have GPUs with their own assembly languages?
> 
> I must confess that I have not kept up with Cg and other recent 
> developments. But I don't consider OpenGL and friends the ultimate 
> graphics solutions. REYES will have real time implementations in the 
> near future (see http://graphics.stanford.edu/papers/reyes-vs-opengl/ 

I read this quickly a while ago... wasn't one of the conclusions that 
there were inherent inefficiencies in the Reyes model in the context of
a hardware implementation?

(feel free to consider that a rhetorical question ;-)

> and 
> http://www.cs.virginia.edu/~gfx/Courses/2002/BigData/lectures/Stream%20Processing/Stream_Processing_6up.pdf

I haven't seen this... I'll check it out.

> ) but I don't think even that bag of tricks will replace ray tracing and 
> radiosity.

One thing about ray-tracing is that it inherently has bad memory
locality (for each pixel, you might have to test for intersection
against each object in your scene).  I don't see a way around this.

I don't know if real-time ray-tracing/radiosity are desirable from the
point of view of amount of computation required.  This is especially
true in the common case of mostly static scenes where radiosity can be
precomputed and stored in textures.  It's my understanding that many
current 1st person shooters use this approach.

> 
> We need better graphics.

I'm always interested in hearing new ideas.  (I'm very excited that I get to
hear a talk on the topic by Pat Hanrahan this Thursday).

Best,
Joshua


> 
> -- Jecel



More information about the Squeak-dev mailing list