NV30 Vertex program performance

[eSa]

Ah, please enlight me gurus :

From "NV30 OpenGL extensions" pdf document :

" Vextex program performance clock for clock

- Over 3x faster than nv20
- over 1.5x faster than nv25

"

So, what kind of performance nv25 (GF4) has ? How many instructions/cycle ? Anyone !?

 

[Saem]

Hopefully I have this right.

GF3 = NV20
GF4 = NV25

I believe the GF4 has two vertex pipes. Though, there might be some resource sharing. I'm not sure about exact numbers, if that's what you wanted.

 

[Xmas]

These numbers suggest three vertex pipelines for NV30, but it also says "over". So that could be four slightly less efficient pipelines.

 

[Dave Baumann]

Or two slightly more efficient.

 

[MfA]

Thats unlikely Dave, they would have to be superscalar (or long instruction word).

Could just be they have 4 with the same efficiency individually as GF4, but with the numbers simply recognizing that performance cant scale perfectly linearly with the number of pipes.

 

[Sage]

it will be 4 bigger and better pipelines, but performance doesnt scale linearly, as MFA said.

 

[GraphixViolence]

Interesting information. Here's an observation:

NV20: 1 vertex engine x 200MHz = 200M instructions/sec
NV25: 2 vertex engines x 300 MHz = 600M instructions /sec
NV30: 3 vertex engines (effective) x 400 MHz (?) = 1200M instructions/sec

If you look at the quoted Mtri/sec figures from Nvidia's documentation, they are:

NV20: 40+ Mtri/sec
NV25: 136 Mtri/sec

These numbers more or less line up with the predicted instructions/sec ratios given above. By extrapolation for NV30, assuming 400 MHz, you would get ~270 Mtri/sec. That's considerably less that the 325 Mtri/sec ATI is already quoting for the R300. Of course, these are all purely theoretical numbers, but interesting nonetheless.