Substantiated NV20 specs

[_Silk_]

It seems those earlier NV20 rumors might be true.

Go here and here to check it out yourself.

It would seem Microsoft released the final specs for the X-Box.
100 million sustained polygons per second
4.8 gigapixels per second
150 million particles per second (huh??)

Now some of those numbers (4.8 Gigapixels) are the same as those rumored from yesterday for the NV20.

Then in the second article we have an interview with J Allard from Microsoft. What he reveals is that the X-Box is not based off the NV25, but actually the NV20. They call the chip the NV2A, indicating that there are some differences between the X-Box chip and the consumer NV20..... but how many.

Now, if you want to take these numbers as BS marketing numbers that the chip will actually never reach or not is up to you. The Microsoft spec didn't mention anything about embedded memory, HSR, or tiling methods. So I guess we have to wait and see if the bandwidth problem has been adequately addressed or not.

Whatever the case, it seems nVidia is making one kick-a$$ chip for this winter.

(Side Note: 4.8 Gigapixels!! !#$$! Is Microsoft sure they didn't mean GigaTexels - if not how many GigaTexels can this beast do....... Better not think along that line. Mabye its not quite Toy Story, but from the sounds of it their getting there in leaps-n-bounds)

 

[Soccerman]

as in the previous thread, 4800 is just PR.. the real rate will probably be around 1200mpixels, with their HSR boosting it up to a theoretical 4800, if the scene has a 4X average overdraw.

I'm guessing that is still the NV25 in there, however it probably has more texture units per pipe then the NV20..

 

[Fozzie]

The XBox GPU is almost certain to be a vanilla NV20 with Memory controller & Northbride funtionality. Probably nothing extra in terms of texture units or whatnot. Looks like about 10 million or so extra transistors compared with the PC NV20.

The rumor mill is that the NV20 is taping out now with final silicon in December. This also matches exactly TSMC's .13 proccess roll-out plans. So it doesn't take a leap of intellect to conclude that the NV20 will be fabbed at .13 running at 300MHz.

Most likely cards shipping in quantity in late Feb, with reviews in mid-late Jan depending on the NDA's. Give or take 2-3 weeks.

 

[_Silk_]

I thought it was the NV25 in the X-Box too until I read the interview at ZdNet with J Allard from Microsoft.

Here is the quote from the interview:
GS: So what are some of the key technical challenges you see to hitting the specs you've released? It's been said that the graphics are based on the Nvidia NV25 chip that's two generations away. With the NV20 said to be slipping a bit, will it be hard to make your schedule?

JA: I can't really go into a ton of detail about the Nvidia relationship and our approach to the engineering of it. But I can say the NV20 is the core of the Xbox graphics. The guys who are working on the NV2A - we call it the NV2A inside of Xbox - are working in parallel with the NV20 folks, so they sort of marry back upafter the NV20 hits the shelves. Our developers should be getting NV20s real soon.

However, JA could just be confused. After all the NV25 is just the Fall (?Summer? was supposed to be Spring, but considering NV20 is late...) refresh of the NV20. So he could still be correct in saying the X-Box is based off the NV20 core. But still... he sounds pretty sure of himself.

Also, you're correct (or appears to be) on the 4800 PR number. Ixbt-Labs info says it is a 300Mhz chip, with four pipelines, which equals 1200Mpix/sec. It has HSR, and then assuming an average screen overdraw of 4, they get there 4800Mpix/Sec. Still will rock tho..

All of the rumors point to this 4800Mpix/sec number - same as the X-Box spec. If these rumors are correct, it would definitely seem the X-Box is based off the NV20 and not NV25.

Now this still leaves the big question - memory bandwidth bottleneck. I know HSR will help this by not rendering unseen polys, and thus reducing Z-Buffer accesses. But I don't think it will be nowhere near enough. By looking at 3dMark 2000 fill rate results, you can see that the GeForce 2 only gets ~1.2Gtex/sec at 16bit color (max. is 1.6Gtex/sec) and significantly less in 32bit color. And by looking at the test, HSR will not help at all as the test is one big poly that fills the screen. It is simply a raw fill rate test - and the GTS does not have enough bandwidth it reach its potential.

Going by the 1200Mpix/sec and assuming two-texture units per pipeline, we have 2400Mtex/sec. With faster DDR memory they might be able to come close in 16bit color (as the GTS did), but 32bit is a different ball game. How is nVidia going to provide enough bandwidth?

Have to wait and see....