only 1TMU

[sebazve]

so i guess with that bandwith is just overkill to have 2tmus

[nutball]

1TMU where?

[StealthHawk]

Originally posted by nutball
1TMU where?

per texture pipe...on NV30

[nutball]

Where does it say that?

What I read was:

Interestingly, neither Kirk nor Tamasi were willing to disclose how many texturing units are present in each pixel pipe. In our interview, they both stressed the decreasing importance of pipes, and focused on the number and sophistication of internal processing units -- which they feel is more important. Kirk describes GeForceFX as a network of processing units on a single die

Am I behind?

I can read the above in two ways: a) they won't talk about because their part isn't great (ie. you're right, 1TMU per pipeline), or b) the second bit is right.

Frankly the way I see GPUs developing it's the latter. The parallel with CPUs is clear: multiple logical processors (pipelines) on top of a different number of physical computational resource units.

[StealthHawk]

the previews i have read(and information in this board) says it has 1.

maybe everybody's wrong, but i doubt it.

[nutball]

Hummmm. The fog deepens...

[Bigus Dickus]

Yes, 8 x 1. When the R300 was introduced, many people tried to explain why it made sense to "go back" to a 1 TMU per pipeline design for modern GPU architectures. I guess now that nV has done the same, people will finally believe it.

[Lezmaka]

What I found odd after reading several previews is that something that was said in one preview, another said "we don't know". The # of tmu's was one, there were a couple others, but can't remember what they were.

[StealthHawk]

each texture unit can already do 16 textures per pass, perhaps this is how they made up for it the lack of an 8X2?

[sebazve]

i guess for newer games the number of tmus doesnt matter it all comes down to vs/ps power:D

[nutball]

Hmmm. I saw the 16-textures-per-pass bit. I wonder if we're talking sampling 16 textures but with 8 unique sets of texture coordinates.

[gstanford]

from the extremetech article:
Interestingly, neither Kirk nor Tamasi were willing to disclose how many texturing units are present in each pixel pipe. In our interview, they both stressed the decreasing importance of pipes, and focused on the number and sophistication of internal processing units -- which they feel is more important. Kirk describes GeForceFX as a network of processing units on a single dye, and adds:


Pipes don't mean as much as they used to. In the [dual-pipeline] TNT2 days you used to be able to do two pixels in one clock if they were single textured, or one dual-textured pixel per pipe in every two clocks, it could operate in either of those two modes. We've now taken that to an extreme. Some things happen at sixteen pixels per clock. Some things happen at eight. Some things happen at four, and a lot of things happen in a bunch of clock cycles four pixels at a time. For instance, if you're doing sixteen textures, it's four pixels per clock, but it takes more than one clock. There are really 32 functional units that can do things in various multiples. We don't have the ability in NV30 to actually draw more than eight pixels per cycle. It's going to be a less meaningful question as we move forward...[GeForceFX] isn't really a texture lookup and blending pipeline with stages and maybe loop back anymore. It's a processor, and texture lookups are decoupled from this hard-wired pipe.

In our discussions with Kirk and Tomasi, we returned to the topic of putting those 32 floating-point processing units to work. The GeForceFX can operate on 16 texels per pixel per rendering pass without having to do multi-pass rendering into the frame-buffer.

This takes multiple clock cycles, but because nearly all this work occurs without touching the frame-buffer, these operations should be carried out quickly. When it's done, the final pixel color is finally written to the color buffer.


Greg

[StealthHawk]

Originally posted by nutball
Where does it say that?

What I read was:



Am I behind?

I can read the above in two ways: a) they won't talk about because their part isn't great (ie. you're right, 1TMU per pipeline), or b) the second bit is right.

Frankly the way I see GPUs developing it's the latter. The parallel with CPUs is clear: multiple logical processors (pipelines) on top of a different number of physical computational resource units.

read the interview at B3D with that guy from nvidia. i think in there he said it was 8x1. the interview was done today....well, yesterday

[nutball]

Yeah, I read a few interviews yesterday which clarified things.

In particular was one guys from NVIDIA who said they sample 8 textures per clock, but can address 16 textures at one time (the implication being the other 8 would be sampled on the next clock).

That makes it all clear now.

Note to self: must stop using more than 16 textures per pass :cool: