New VPU's

[Inspirer]

Seeing the Radeon 9700 Benchmarks really blew me away,
but then I came to think about the overkill that is required to get these results using conventional immediate mode rendering.
Now even more so than before, because pixels can now be shaded using very long programs only to be covered up by some other pixel which might also be covered.
Add AA + AF to the mix which are also being done on hidden pixels.
And don?t forget that every pixels can now take up 128 bit of FP data.
The impact of overdraw is HUGE !!

The early Z tests are far from entirely eliminating overdraw because it all depends on the order in which the application is sending polys to the card.

If a 110M transistor chip that suffers from overdraw can handle 1024x768 + x4 AA + x16 AF, what kind of performance could we expect from a deferred, tile based rendering chip with 8 pixel pipelines?

Or, on the flip-side, how much would a deferred, tile based rendering chip with equal performance and feature set cost?

Look at this Roundup, look at Kyro?s benchmarks multiply them by 4 (kyro w/ 8 pipes) then multiply the result by 1.85 (running at 325MHz instead of 175).

The results exceed those of the Radeon 9700 !!!

 

[Wolfsraider]

so let me get this straight

you are saying that a kyro with those performance enhancements would beat all the rest.
yes i'd love to see it.
but the truth is it is all speculative
the actual numbers may be lower than projected (matrox?)

Look at this Roundup, look at Kyro?s benchmarks multiply them by 4 (kyro w/ 8 pipes) then multiply the result by 1.85 (running at 325MHz instead of 175).

again here we see that you are quadrupling the known benchmarks comparing these figures to a set standard .problem is would that give a "true" quadrupling of power in the benchmarks?

then you are almost doubling the clock of the kyro. sure it's all possible at least in speculation and i'd love to see this but...

again until someone makes a card with these "enhancements" who's to say what it will do in real applications?

hope this helps
mike

 

[Inspirer]

As I understand the architecture,
kyro's benchmarks are "safer" to multiply since it has no overdraw.

and, to my best understanding, Kyro's benchmarks under UT2003 expose weaknesses of the card that are relatively easy to solve:

1. memory size. Kyro's 64MB might not be enough with maximum details which leads to texture thrashing.
adding more memory to the card sounds easy enough.

2. the card's pitiful bandwidth kicks in.
again, support for DDR Memory doesn't sound like a major change.

So, in fact, unlike the Pharhelia, I believe that Kyro, given more pipelines, higher clock and more bandwidth
would more closely live up to it's theoretical numbers and have better results in the mentioned benchmarks than simple linear multiplication.

In addition, Kyro's 15M transistors *4 = 60M.
I doubt that a 60M transistor part on a 0.15 Micron Process would have problems hitting 325MHz

But also, you missed my point.
I wouldn't want to see a Kyro (which had no programmable pixel effects capability at all and no HW T&L whatsoever) being enhanced.
I was just using its benchmarks as "proof" that it's architecture is potentially faster.

What I would like to finally see is a high end chip built from the ground-up to both have a complete feature set, and use deferred, tile based rendering. Because I believe it would have a better price/performance ratio.