Nv40 16 full pipelines- The Inq.

Complexity can result in inefficiency, but it certainly is not "an elegant way to avoid talking about inefficiency."

 

The number of pipelines really has nothing to do with complexity. Is the NV30 more complex than the NV31 or NV34? I'd claim that it is not. More specifically, I claim that each individual pipeline of the NV3x processors is more complex than each individual pipeline of the R3xx (one reason, for instance, that the 4 pipelines of the NV35 can compete pretty well with the 8 pipelines of the R3xx).

I was stating that the complexity of the NV3x design, which I claim is certainly more complex than the R3xx design (they try to do more per pipeline...I call that more complex), is one reason for the inefficiencies of the NV3x design. nVidia obviously has to combat this inefficiency for their next major design. The question I was asking was how that inefficiency will be combatted.

Will nVidia drastically change the structure of the pipelines to make them simpler, and thus easier to make efficient? Or will nVidia attempt to fix the pipelines that they have already designed, while at the same time making them more capable? I don't know the answer, and I really don't know which one would be the better choice. But I think it will be interesting to find out what nVidia does do.

Side note:
Another hint to the complexity of the NV3x design is that shader performance has increased dramatically with later shader compilers. ATI's shader performance has changed very little. I claim this is due to the complexity of nVidia's design (this aspect I don't consider necessarily bad for nVidia...it's just more work for their software team).

 

no, Chal, I'd think that is because the compiler is optimizing for register usage (you know, what ps_2_a would be doing if people actually used it ever). if everyone was using ps_2_a right now, we would see absolutely no benefit from NV's shader compiler.

didn't somebody test Shadermark or RightMark with ps_2_0 and ps_2_a and find no difference (or <3%)? that's the compiler for you. it's a case of ps_2_0 not generating optimal NV3x code but generating very good R3x0 code, which is why NV's shader compiler appeared to give such a large performance boost relative to ATI's.

(oh, here we go:

http://www.beyond3d.com/forum/viewtopic.php?t=8623&highlight=ps2a

1 or 2 FPS here and there between ps_2_0 and ps_2_a.)

 

you could always say the improved performance could be because their compile has just been god awful this whole time... it goes both ways. But to me it seems more like app detection.

Everytime they release a new driver with an "improved compiler" it always effects the most benchmarked apps and all the rest of the games still perform like ass.

 

We've been through this before, but it is simply not the case that ps_2_a compilation can be fixed in FXC and everything will be hunky dory. The driver will still have to do significant work, just like the ATI driver today still has to do further work. The pixel shader assembly language does not map directly to the underlying hardware, and does not represent the original semantics of the source properly. See the discussion of "when to choose predicates vs branches vs ..." in the coding group. if the R420 support PS_2_A, do you think FXC can generate one set of code with the PS_2_A profile that is optimal for both PS_2_A compliant cards (NV30 and R420?)

 

Early drivers didn't even have a compiler. Probably a little of both though.

 

Yes the newest RightMark build has PS 2.a support and it makes no difference at all -- at least here.

 

Thanks.

SOON? It's always SOON! incurable doesn't want it SOON! incurable wants it now!!! [/crybaby_mode]

cu

incurable

 

I don't know why you guys are assuming R420 willbe 8-pipe. I hope it's not from that lame article on Anandtech - if Anand had got his info from ATI he wouldn't have been able to use it.

Remember that by the time R420 appears it will have been over a year and a half since R300 appeared, and you think in all that time the best ATI could do was make the pipelines faster?

One more thing to think about - what if R400 was canned not because i wasn't working out, but to release the second iteration of that part as the first chip. Think R350 taking the place of R300.

 

Every reliable source I can think of has said that R420 will have eight pipelines, so I see no evidence to the contrary right now.

Besides which, who says that eight R420 pipelines will be the same as eight R300 pipelines?

 

Firstly, I don't think anyone reliable has said that R420 will definitely have 8 pipes. Probably the most reliable source of info around here, Dave B., had this to say in response to the Anandtech story referencing the 8-pipe config:
"Salt, pinch of"

Secondly, no-one said that if R420 had 8 piplines they'd be the same as R3XX. I certainly didn't, anyway.

 

Well, most obviously with the AT story was that the fill-rates just don't add up.

 

Has anyone confirmed the 6 vertex units?

 

Of course they had compilers, since day one. You can't run DXG tokens directly on hardware, irrespective of who built it. You can't even run DX6-style multitexture setups on hardware without translating them into something.

You've been fooled, I'm afraid. Compilers have been with us for ages. Optimizing compilers, and good ones, well, that's a different story. That really just started with DX7 and the Geforce's register combiners.

 

Dave has been hinting otherwise all over the forums here...

 

First, I don't think that's true.

Second, PS_2_a didn't come out until long after the NV3x was first available. Why do you think that was? It certainly wasn't because the NV3x was easy to optimize for.

 

So what's the speculation on ATI's pipelines? I heard conjecture a couple of months back on 12x1. Would this be 12 true pipelines, or maybe 12 extreme pipe?

 

cant say for anybody else but mufu says 8 with some extreemely beefed up shaders of some kind and hellbinder says it will be more than 8 pipes. Since both of these guys have some kind of info that is often very reliable I guess we will just have to wait and see.

Another note is that helbinder has been saying 12 or 16 pipes for the r420 for quite a while now. Maybe he knows something the rest dont know this time around. Another note Hellbinder seems to know people in the rd or something simmilar and as such he knows about products they could release but dont always make it to market.

 

Maybe it will be the R420 that will be able to access the VS units making the 8 PS extreme pipes able to run as 12x1...

 

The common belief is 8 "extreme" pipelines. Hellbinder has hinted at 8 pipelines operating as 16 in some cases. Both of those rumours could actually be identical, or maybe they're different. It wouldn't suprise me if they were both wrong.

Or not. Or yes. Something.