![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
INVESTIGATIVE JOURNALISM is not an easy thing. But we do what we have to do just as Nvidia and ATI need to make their chips. Our job is to post news as exclusive as possible, this is occasionally at the risk of jeopardising communications with the company that we write about. Still we have to do it.
Anyhow, we have something that you all have been waiting for.
It was easy enough to state that Nvidia's NV40 was an eight-times-two pipeline architecture before and now claim it's 16x1, but it was very hard to find out how this chip actually works without being able to speak with the company's David Kirk and ask him all the nasty questions. We dont sign NDAs, thats our problem. If we did, we'd know, but we wouldn't be able to tell you. As it is, we can post this stuff as soon as we get it.
So, the claim that we made more than once that NV40 is 16x1 chip is partially true and but the claim that I made previously that the chip is 8x2 is also true. How's that possible?
NV40 will score magnificent 6500 Mtexels per second in 3Dmark01 whereas ATI's Radeon 9800PRO will give you around half that, but there is a catch here.
We understand that the NV40 will actually only have eight physical pipelines, but these will appear act like 16 in certain games and, indeed, in 3DMark 2001 Nvidia was telling people how 3DMark01 is a very nice benchmark since they can render 16 textures per pass in it, using only eight pipelines. What Nvidia is using is the ability of the Vertex Shader model 3.0 (known as PS 3.0 or VS 3.0) where the Shader can actually render textures as a virtual pipeline but can only render them without filtering information.
But, in modern games you always apply Bilinear, Trilinear or Anisotropic filtering to textures that you render, otherwise they will have some visual malformations. This is why, in normal games, eight pipelines will be able to render eight textures in single texturing pass and possibly even 16 if you render both textures in same pass using an 8x2 architecture.
What we understand Nvidia is doing is processing eight textures by pipelines while the other eight textures get processed by the Vertex Shader 3.0. If they decide to use this in modern games this would roughly mean that they can render every second texture properly, which would mean that you would be able to notice the difference. However, this can be tweaked in the driver if they decide to go that way, but image quality would have to be sacrificed for that.
To get into more details, the Vertex Shader model 3.0 feature has the ability to fetch textures but you could use your VS 3.0 only where you dont have any geometry - when you have to process just pixels. This cannot be used in modern games as you process just point sample textures. In 3DM2001SE you get such high numbers because you deliver 16 pixels per clock.
We asked Nvidia to comment but since they havent announced NV40 they sdaid they can't comment on "unannounced products". But we expect they will be able to tell us more then they launch on 14th in the old continent.
ATI's R420 offering remains at 12x1 architecture which makes this games-running business very complicated, since you cannot bring any final judgment about who will win the chip speed race.
However, we expect NV40 to deliver 16x1-like performance in Doom 3 (or should we say 16x0, which is how people call Nvidia's approach) where they render all the information except Z or colour features. This will be used by Doom III and games based on this engine.
It's likely that Nvidia will launch up to two weeks prior to R420's introduction. µ
http://www.theinquirer.net/?article=14878
hmmmm k, i think this is bad
Anyhow, we have something that you all have been waiting for.
It was easy enough to state that Nvidia's NV40 was an eight-times-two pipeline architecture before and now claim it's 16x1, but it was very hard to find out how this chip actually works without being able to speak with the company's David Kirk and ask him all the nasty questions. We dont sign NDAs, thats our problem. If we did, we'd know, but we wouldn't be able to tell you. As it is, we can post this stuff as soon as we get it.
So, the claim that we made more than once that NV40 is 16x1 chip is partially true and but the claim that I made previously that the chip is 8x2 is also true. How's that possible?
NV40 will score magnificent 6500 Mtexels per second in 3Dmark01 whereas ATI's Radeon 9800PRO will give you around half that, but there is a catch here.
We understand that the NV40 will actually only have eight physical pipelines, but these will appear act like 16 in certain games and, indeed, in 3DMark 2001 Nvidia was telling people how 3DMark01 is a very nice benchmark since they can render 16 textures per pass in it, using only eight pipelines. What Nvidia is using is the ability of the Vertex Shader model 3.0 (known as PS 3.0 or VS 3.0) where the Shader can actually render textures as a virtual pipeline but can only render them without filtering information.
But, in modern games you always apply Bilinear, Trilinear or Anisotropic filtering to textures that you render, otherwise they will have some visual malformations. This is why, in normal games, eight pipelines will be able to render eight textures in single texturing pass and possibly even 16 if you render both textures in same pass using an 8x2 architecture.
What we understand Nvidia is doing is processing eight textures by pipelines while the other eight textures get processed by the Vertex Shader 3.0. If they decide to use this in modern games this would roughly mean that they can render every second texture properly, which would mean that you would be able to notice the difference. However, this can be tweaked in the driver if they decide to go that way, but image quality would have to be sacrificed for that.
To get into more details, the Vertex Shader model 3.0 feature has the ability to fetch textures but you could use your VS 3.0 only where you dont have any geometry - when you have to process just pixels. This cannot be used in modern games as you process just point sample textures. In 3DM2001SE you get such high numbers because you deliver 16 pixels per clock.
We asked Nvidia to comment but since they havent announced NV40 they sdaid they can't comment on "unannounced products". But we expect they will be able to tell us more then they launch on 14th in the old continent.
ATI's R420 offering remains at 12x1 architecture which makes this games-running business very complicated, since you cannot bring any final judgment about who will win the chip speed race.
However, we expect NV40 to deliver 16x1-like performance in Doom 3 (or should we say 16x0, which is how people call Nvidia's approach) where they render all the information except Z or colour features. This will be used by Doom III and games based on this engine.
It's likely that Nvidia will launch up to two weeks prior to R420's introduction. µ
http://www.theinquirer.net/?article=14878
hmmmm k, i think this is bad
...
