NV30 has 125 million transistors

I find the "Just Like The Movies" statement rather silly

 

anyone took a look of that slide presenting nVidia's technological steps before NV30?? am I the only one that thinks there's definately something wrong on NV25 launch time on that list?

 

What they've done in the past ("Past results are no guarantee of future performance") is to compare to the previous generation at launch, ignoring any optimization and driver improvements in the interim. This of course gives them a bigger marketing delta even if it isn't there in the real world of what you can buy on the shelf and install right then. By that procedure they would be comparting to a TI4600 at launch of the GF4.

 

yeah, NV25 (GeForce 4 Ti) came out in Feb-April 2002. Although I think it was
ready in fall 2001. it had taped out. The XBox GPU, NV2A, which taped out in Feb 2001 is similar to NV25 in mosts respects, though NV25 has quite a few additions and is a higher performance part.

 

I wonder if 8 pipelines will be the standard for the NV4X series, as it will be with the NV3X series. 4 pipelines was stardard from NV10 through NV25. If the size and performance of Nv processors keeps increasing rapidly due to increased competition from ATI, perhaps we'll see 16 pipelines with NV4X, but if not, then probably with NV5X and XBox2.

 

Well, 51 billion, divided by 400 million (clockspeed) and 8 (number of pipelines in pixel shader) yields approximately 16. That means 16 flops per pipeline per cycle. We know the pixel shader is fully floating point. The pixel processing unit can execute 2 half float ops per cycle (assuming scalar and vector in parallel, it would be 4 flops), there are two tmu's which may also be capable of two half float ops per cycle (that is 2 tmu's times 2 which gives 4 flops), and a texture address unit which must provide the tmu pipes and pixel program processor with texture data (4 half floats for the tmu's and 2 for the pixel program processor, for a total of 6 flops). This adds to 14 flops and is the closest estimate I could give for the 16 flops per cycle performance of the NV30. Does anyone agree/disagree?

If this info holds true, the NV30 is one impressively parallel beast.

Edit: Oops, I forgot to note that vector ops count as 4 flops (assuming simd with 16 bit half floats). This would mean 10 flops for the pixel processing unit, 4 flops for the 2 tmu's, which would leave 2 flops (2 half floats per cycle) for the texture addressing unit, adding up to 16 flops per pipeline, per cycle. This sounds very impressive. If this is the case, the NV30 pixel shader architecture is more ambitious than even the R300's 3 issue pixel engine. The R300 can execute 3 instructions per pipeline (128-bit) and the NV30 seems to be capable of 8 (6 if not counting vector and scalar ops separately at 128-bit).

 

22-bit post filter?

 

If you haven't read my last post, read it, I'm wondering if it seems to add up correctly. Secondly, I would like to know if one could assume that the NV30 would be capable of executing quarter floats (8-bit) in a quarter of the time it executed full floats and half the time it executes half-floats?

 

Well, Nvidia said at E3 2001 that the fall product was more than a Geforce3, so you would assume they were talking about Geforce4 and not Geforce3 Ti500/200. And Anand , indirectly , at least , confirmed this in his Ti500 review. 2H 2001 was when I expected Geforce4 and Parhelia which would have been far more exciting than G550 and Ti500 versus 8500 but ah well.

When I asked Tony on his E3 2001 comment at E3 2002, his answer was interesting though, I'm not sure I can post it.

 

Marketting always exagerrates things, so it's guaranteed the NV30 will not be literally 2x faster than GF4. If it was then it'd be "4 times faster", because they'd use some bizarre situation in which it really was that much faster as a basis for it.

 

Hehe I can create situations where my 9700 Pro is up to 8 times faster than my Ti4600 (probably even more if I want to toy around with some settings)

FableMark using FSAA and Aniso kills the Ti4600 while the 9700 Pro strolls along without noticing anything out of the ordinary.

So I guess you could say that not only is the 9700 8+ times faster, it also has a 8+ times more efficient HSR-method.
But then again, we all know that is a load of crap.

When using FSAA and Aniso the 9700 Pro is 2-3x times faster than a Ti4600 and that's no BS, I'm assuming Kirk is talking about a similair situation if so then it should pretty much be on par with the 9700...

I mean there's even "normal" extreme gaming situations where the 9700 pro is 5 times faster. Check anandtechs results in JKII running at 1600x1200 with 4x FSAA and 16x Aniso where the 9700 is exactly 5 times faster (mind you that the Ti4600 uses half the "amount" of aniso so it's it's even better than 5 times faster).
And it's not like "oh well it's unplayable on both cards anyway" the difference here was 20 vs 100 fps.

I dunno, I'm just worried about FSAA and Aniso performance on the nv30.

 

hehe the GF4 was also supposed to hand us FF-realtime, hmm now if the all-mighty 9700 pro can't even do it (well I assume that what they want to imply by saying that the nv30 will be able to) then what could a poor old Ti4600 possibly do

 

Especially about aniso: none of those slides mentioned...

 

Well, if anyone still have any doubt:

http://www.nvidia.com/content/areyouready/facts.html

 

you mean it can speak both french and german??

 

"If converted to US Dollar bills and placed end-to-end, would stretch from New York to Hong Kong."


hopefully they are not reffering to the cost of the card here...

 

LOL!