[Fwd: Re: [Open-graphics] Trading off time and space]

[Nicolai Haehnle]

On Monday 06 December 2004 13:45, nico at seul.org wrote:
> > Anyway, I'd still be very interested in comparing gate count for current
> existing graphics cards to what the FPGA can offer. According to
> http://www.hwhunpage.hu/gvt_e.htm, a very modest TNT2 chip takes 8
> million transistors. It would seem like a good starting point if we can
> fit that onto the FPGA. Anyone have some good numbers of how this maps?

Always keep in mind that the TNT2 has two complete texture units in hardware 
(at least as far as I know). This includes all the logic for mipmap 
selection and for filtering, which is quite expensive especially in terms 
of multipliers etc.

It might be a valuable feature-and-space-vs.-performance tradeoff to reuse a 
single texture unit for multiple textures (by loading different memory 
offsets and texture specifications) when multi-texturing is needed. This is 
bound to buy a *lot* in terms of flexibility (how else would you put enough 
texture units for interesting shader work onto the chip?), and maybe it's 
even possible to do this without hurting the single-texturing case (which 
is important for desktop work).

As always, I'm not sure if this is realistic, since I'm not a hardware 
person. I just wanted to point out that we're in no way forced to follow 
the same design decision as NVidia back then. There are many creative ways 
in which the transistor/gate/whatever budget can be used.

cu,
Nicolai
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://lists.duskglow.com/open-graphics/attachments/20041206/ee36a71d/attachment.bin