R520: 32 PS pipelines? Any truth to this?

Whats bothering me is (if those dreadful rumors are true of course) that R520 was originally an 32 pipe design but ended up with 16 pipes because of some yield issues... wouldn't be good for Atis reputation

 

Is there bad news tho? Maybe you're right. <shrugs>

What I read was NV sources starting to run from the 24/32 rumor and trying to build a plausible exit ramp to make their escape on, and still getting a kick to the ribs in at ATI on their way out the door.

But, y'know, maybe that's just me.

 

Re: WTF does that mean? (I'm going nuts!)

FUDDetect -post nagus

*program crashes followed by a BSOD*

 

Oh interesting. I hadn't thought of that because I don't have any evidence that NV sources started the 24/32 rumour in the first place. I don't see how you could run away from such a rumour though. It should become glaringly obvious whether or not it is a native 16-pipe or "failed" 32-pipe part once it's released. So NV is behind the scenes pushing this nonsense then they'll be outed no matter what they say now.

 

No thats totally not the case...

The whole "pipeline" issue is a misnomer. People are still thinking pixel pipeline wth X number of TMU's and some "shader stufF" tacked on the side.

the new GPu's are "Shader processors" with some "texture stuff" on the side or as an additional function. One texture with multiple loopbacks for multitexture for older games. No one is making games today using 4 textures for a surface. Its all going shaders, Normal maps etc..

16 pipelines or 32 pipelines (pick a number) etc.. can be all true at the same time on the G70 and the R520. It all depends on if its doing a texture, a shader or vertex.

Information can leak that it is a 16 "pipeline" a 32 "pipeline" or heck soon a 64 "pipeline" and in once sense or another they can all be true at the same time. Its comes down to what type of pixel processing is going on at any given time.

 

No one is saying so in public, that I've seen. There have been some "knowing looks" exchanged hereabouts in some posts by some of our beloved wise guys. Again, I like VR Zone, but this coming from them is. . .uh. . supportive of that theory, shall we say, IMHO.

 

Dean Sekulic from Dave's interview - "But, as I said before – the shader that really raped high-end GPU was the one with 8 texture fetches (and only 25 instructions!)."

 

Ok so there is at least *ONE* Guy out there going gonzo on textures...

8 texture fetches.. *whew* the modern GPu's are not gearded to do something like that at all.. what 8 loops?? thats got to create some serious pipeline stalls to..

 

Which theory? That they're a mouthpiece for Nvidia or that the report is FUD because those two are independent.

 

Yeah, but how many of those 8 texture fetches are actually straight-forward surface texturing?

Jawed

 

Does it matter what the samples are used for?

 

I think it does, because I suspect Croteam's algorithms are indicative of what other developers could do - use textures intensively for tasks other than surface texturing.

Jawed

 

I'm pretty sure that Half-Life 2 uses at least 4 textures on most surfaces.

1. Colour-map
2. Normal-map
3. Cubemap (for specular)
4. Specular-map
5. Lightmap

Of course, the spec-map is probably packed into the 4th component of the normal-map, but there may be multiple lightmaps.

 

Isn't D3 taking up to 5 texture fetches in some cases?

 

How many textures are used per shader it's not that important, we should focus our attention to the 'pure' math instructions/texture fetch instructions mean ratio.

 

"Mouthpiece" is a much harder word than I would use. My assessment of their main sources is my own, and yes on the whole over time it seems to me their NV camp sources seem to be much better.

I mean, yield numbers?? Not even a range --spot on number. Who would be in a position to provide something like that with enuf credibility that they'd be willing to use it?

So, if you accept that is an NV camp source piece, and that R520 is and always has been a 16 ps part, the rest follows. I mean, c'mon, that piece screams "oh s*it, it is getting close to when people are going to absolutely know that the R520 sold to the public is 16 ps pipes --how do we wiggle out of our previous misinformation?"

Otoh, if R520 is released as 24 or 32 ps in Aug/Sep, I'll publicly apologize on the B3D Review thread to both VR Zone and NV. But, you see, from their own piece that is unlikely to be necessary. . .

 

Well it seems to me that with algorithms like precomputed radiance transfer there's no short-term escape from a greater demand for texture-based algorithms.

But hey, I'm no developer. Just observing from the sidelines.

Jawed

 

Geo, the bit you might have trouble wiggling out from (if my sig's theory is true) is that some people will call R520 a 32-pixel pipe even though it isn't...

I suspect the qualifier "fragment pipeline" is worth being boringly specific about

Jawed

 

Which is exactly why having a shed load of TMUs isn't that important, yes? I mean if one has a pixel shader with a arithmetic-to-texture op ratio of, say, 10:1 then the TMUs are going to be idle for quite a few clock cycles. More than enough time to fetch texels; Xenos can handle up to 12 2x2 pixel quads at a time and yet makes do with "just" 16 texture samplers.

 

Hmm! Good point. So revised. Wow, imagine what R520 is going to cost when they are getting much less than 3% yield for that "32-piper" One die per wafer?