A few questions on NV30, NV35

[antlers]

In the past, nVidia has gone to some lengths to avoid putting an extra power connector on their cards, in part perhaps to make their technology more palatable to OEMs. The original GeForce 256s were at the very edge of compliance with the AGP spec (the really should have had an extra power connector), and the GeForce 4 4x00 suck every bit of current they can get out of the bus.

It's not unreasonable to suppose that when they planned the NV30 nVidia thought they would get market-crushing performance by staying within the AGP spec at .13 microns, and only later learned that the performance they envisioned at the time the chip was in it's early planning stages wasn't so far out of the reach of their competitors. You would not say that their design was limited by target power consumption so much as it was limited by their original target clock speed, which was chosen on the basis of a lot of factors, including power consumption.

 

[RussSchultz]

I guess theres something to be said about efficient design

gf4 (60m?) > parhelia (80m)

Parhelia is more feature rich, hence more transistors.


Anyways, transistor count is generally no indication of speed.

 

[alexsok]

I guess theres something to be said about efficient design

gf4 (60m?) > parhelia (80m)

That puzzles me too...

Intresting conclusion from Digit-Life about this:

The performance of vertex shaders with the promised 4 pipelines is too low. I think that only one of them really works. It's possible that lack of the DirectX 9.0, and therefore, the version 2.0 could cause locking of 3-4 pipelines. Next time we will check this hypothesis.

The whole review here:
http://www.digit-life.com/articles/parhelia/index.html

 

[RussSchultz]

Reviewing nVidia's history is not speculation.

No, but drawing conclusions from it is. Speculation is taking past performance and assuming it equals future performance.

If we want to do that, then lets stroll down history lane for everybody:

TSMC has good yields
ATI has crappy drivers
NVIDIA's drivers r0x0rs

See how pointless that is?

(Of course, its probably not as pointless as continuing this discussion)

 

[ERP]

I think the Digit Life conclusion is bogus. VS2.0 doesn't add anything taht would magically allow you to use more pipes.
It's more likely that the issue is with instruction latency in the vertex shaders.
Nvidia basically runs multiple vertices in parrallel through each of the shaders to hide instruction latency, you can turn this off on the Xbox devkits and the performance drops like a stone.
This is obviously jyst speculation on my part and if this is the case Matrox might be able to better utilise the pipelines in later drivers by rearranging instructions in the CompileShader call.

Although I don't doubt that there are issues in Matroxes driver that prevent parhelia from performing aswell as it can (especially in CPU limited instances), my guess is that the issues with Parhelia are more likely related to a combination of cache performance, and probably an over reliance on the raw bandwidth advantage they have.

 

[pascal]

You forgot one, I also had to speculate that the NVIDIA engineers didn't transform into monkeys overnight.

Hehe.

I agree.

The thing that Joe and Pascal want to say is:

what if:

1: Nvidias engineers will be really careful about meeting the AGP specs even though they might have to lower their specs because of it. Even though it might be the difference between 10 % faster or 10% slower (equals to a huge difference in this business)

2: They are getting lazy and won't handtune their design since they don't "have to".

3: A extra power connector looks really bad when trying to market the card at OEM's. Thus avoid it even though it .. (see 1)

Then, the advantage of the 0.13 micron process might not be an advantage anymore.

Bjorn

Dont try to put words in my mouth.

I have been as much technical as I can. Dont turn thinks into some kind of fan dispute because I am nobody fan.

Again, again and again:
1- The R300 is using extra power beyond AGP specs. From the performance standpoint it is an advantage. And you cannot negate that. How much this is an advantage I dont know but maybe the stable under spec R300 work at 20% lower frequency.
2- The R300 design probably is a very good design from the microarchitecture and vlsi design standpoint. (Again I am not saying that NV30 has a bad design)
3- About the OEM maybe they will want cards that work under AGP spec (and maybe is maybe, to be sure contact them).

About the metamorphosis you (edited: I mean Russ) described from nvidia engineers to monkeys it is something that never crossed my mind

We dont know many things about NV30 to speculate about its performance this is all I say. I am not saying it will be faster, or much faster or slower (from the sustained performance viewpoint) because we dont know many things: is it using some kind of deffered rendering, 128bits, any special trick, is the additional logic slowing it, how good is the architecture, how good is the microarchitecture and vlsi design?

You may just use your fan´s faith and believe it is faster. You can believe in the rumours about 48GB/s. You can believe in many things but from what we officially know (120M transistors with .13 micron process) we cant conclude it will be much faster.

edited: english again

 

[Joe DeFuria]

Sigh...

No, but drawing conclusions from it is. Speculation is taking past performance and assuming it equals future performance.

I am NOT assuming past=future. I am saying based on the past, it is REASONABLE TO SUGGEST that the future would be similar.

Russ...YOU are the one who is closer to "drawing a conclusion" than I am. That's the point. You don't see anything that would even SUGGEST that Nv30 clock speed might be lower. That MEANS that you have already CONCLUDED certain things, for example, about those basis I suggested.

FOR EXAMPLE:

1) YOU are saying that "based on the past, it is NOT REASONABLE TO SUGGEST the future will be similar."

2) YOU are saying that "based on the known 0.13 TSMC yield issues, it is NOT REASONABLE TO SUGGEST that it can impact shipping clock speeds relative to design target.

Do you REALLY think those SUGGESTIONS are UNREASONABLE?

 

[8ender]

I think we'll have to wait and see,

as soon as the availible data is exausted, any good conversation will degenerate into conjecture.

Hell, look at Anand's article, even he is guilty of posting rumors as fact.

We have no idea what clock speed the NV30 is going to debut at and yet his article finalizes by saying the NV30 will have superior clock speed.

Either he knows something we dont or he's been fed a line.

I think its time to step back and say "lets wait and see" instead of trying to bury each others opinions on a piece of silicon that isnt likely to see the light of day until december.

 

[pascal]

But it is fun

 

[martrox]

But it is fun

It is? Oh, wait, you're right, IT is fun.....

 

[Nagorak]

Let me state very clearly: I'm not saying that NV30 will be faster, I'm stating that there's nothing out there to suggest that it won't be (other than speculation).

And there's nothing to suggest it will be besides speculation, so where does that leave us?

Bottomline:

1) Everyone expected the R9700 to run at 250 MHz or less. People suggested it was impossible for a 0.15 chip of that complexity to run at 325 MHz.

2) No one ever suggested the NV30 would be almost 200% as fast as the 250 MHz Radeon and run at 400 MHz. Everyone's expectations were that the NV30 would run around 300 MHz, with the process shrink basically negating the increased complexity.

3) ATi pulls a rabbit out of their hat and gets their part running at 325 MHz. That was an unexpected surprise.

Now everyone expects that NV30 will run at 400 MHz just because ATi got their chip running at 325 MHz? All the while they're getting yields of 15%.

 

[RussSchultz]

All I can say is you guys are totally full of shit....saying that there's nothing to suggest it won't be (significantly) faster is just ridiculous.

Yeah, thats exactly what I said. I am so full of shit, aren't I?

Did you even read the whole thread? I hate to be combative here, but constantly CONSTANTLY people don't read half of what's written, and then invent the other half. I said nothing of the sort.

 

[Bjorn]

1- The R300 is using extra power beyond AGP specs. From the performance standpoint it is an advantage. And you cannot negate that. How much this is an advantage I dont know but maybe the stable under spec R300 work at 20% lower frequency.

And again, it just that we're seeing things a little bit different.
I see it as something they have to do because of the 0.15 microns process. You seem to compare it with adding a Turbo to a engine.

2- The R300 design probably is a very good design from the microarchitecture and vlsi design standpoint. (Again I am not saying that NV30 has a bad design)

Ok, but if you're not saying that the NV30 has a bad (maybe not bad but at least not as finetuned as ATi's) design, how can you
say that this is a advantage for the R300 ?
What we do know though is that the NV30 uses 0.13 micron vs 0.15 for the R300.

3- About the OEM maybe they will want cards that work under AGP spec (and maybe is maybe, to be sure contact them).

Maybe they will. But what we're talking about here is if Nvidia would risk their performance crown (when the NV30 is released )
just to keep within the AGP power specs.

You may just use your fan´s faith and believe it is faster. You can believe in the rumours about 48GB/s. You can believe in many things but from what we officially know (120M transistors with .13 micron process) we cant conclude it will be much faster.

I wouldn't bring in any "fan's faith" and stuff like that in this discussion. Especially since i haven't said that i believe in any of those things.
And yet again (yes Russ, i share your tears ), NOBODY has made any conclusions here. Do you mind to quote any of these supposed conclusions ?

 

[mboeller]

All I can say... ...is just ridiculous.

Can You please edit out this sentence? Why do You disqualify Your arguments with such an idiotic statement.

thanks

 

[pascal]

1- The R300 is using extra power beyond AGP specs. From the performance standpoint it is an advantage. And you cannot negate that. How much this is an advantage I dont know but maybe the stable under spec R300 work at 20% lower frequency.

And again, it just that we're seeing things a little bit different.
I see it as something they have to do because of the 0.15 microns process. You seem to compare it with adding a Turbo to a engine.

But without this extra power probably they would have to lower the core frequency to something like 250MHz then it really help. Probably this thing is generate ~70% more heat than the GF4Ti4600

2- The R300 design probably is a very good design from the microarchitecture and vlsi design standpoint. (Again I am not saying that NV30 has a bad design)

Ok, but if you're not saying that the NV30 has a bad (maybe not bad but at least not as finetuned as ATi's) design, how can you
say that this is a advantage for the R300 ?
What we do know though is that the NV30 uses 0.13 micron vs 0.15 for the R300.

What I was trying to say is that we cannot expect the NV30 to have microarchitecture design advantages, maybe it has some graphics architecture advantages we dont know.

3- About the OEM maybe they will want cards that work under AGP spec (and maybe is maybe, to be sure contact them).

Maybe they will. But what we're talking about here is if Nvidia would risk their performance crown (when the NV30 is released )
just to keep within the AGP power specs.

We never know.

You may just use your fan´s faith and believe it is faster. You can believe in the rumours about 48GB/s. You can believe in many things but from what we officially know (120M transistors with .13 micron process) we cant conclude it will be much faster.

I wouldn't bring in any "fan's faith" and stuff like that in this discussion. Especially since i haven't said that i believe in any of those things.
And yet again (yes Russ, i share your tears ), NOBODY has made any conclusions here. Do you mind to quote any of these supposed conclusions ?

Then what are we discussing?

 

[8ender]

1- The R300 is using extra power beyond AGP specs. From the performance standpoint it is an advantage. And you cannot negate that. How much this is an advantage I dont know but maybe the stable under spec R300 work at 20% lower frequency.

And again, it just that we're seeing things a little bit different.
I see it as something they have to do because of the 0.15 microns process. You seem to compare it with adding a Turbo to a engine.

But without this extra power probably they would have to lower the core frequency to something like 250MHz then it really help. Probably this thing is generate ~70% more heat than the GF4Ti4600

Maybe I'm wrong, but isnt more heat produced by higher voltages and not higher clockspeeds?

I seem to remember something from overlclocking CPU's to this effect.

 

[pascal]

More heat can be produced because of one or more factors below:
-vcore voltage (1v, 1.2v, 1.5v, etc...)
-core frequency (200MHz, 300MHz, etc...)
-fabrication process (.15 micron, .13 micron, etc..)
-number of transistors (63M, 107M, 120M)
-pattern of transistors activity
-IO bus activity. (RDR, SDR, DDR, 256bits, 128Bits, 64bits, etc...)

IIRC the R300 has more/less the same core frequency as the GF4Ti4600, the same fabrication process (TSMC .15 micron), probably the same vcore (guess) but has 70% more transistors (107/63=1.7) and twice the IO activity (256bits memory bus). I suppose the pattern activity is not much different ( a lot of parallel activity), then some guesstimate of 50~70% more power consumption.

But see it is a guess.

 

[Dio]

pascal covers this and more, but I thought I'd add a touch more detail:

IIRC, heat is proportional to the square of the voltage and to the total number of transistor state changes per unit time (therefore, it is proportional to frequency).

This is why clock gating (not clocking any unused part of the chip) is a big thing in mobile parts.

 

[8ender]

aah thank you, perhaps what I knew was an urban myth of sorts

but it is related to power consumption correct?

perhaps I was thinking of CPU's wherein the voltage is the main factor affecting heat output.

 

[RussSchultz]

Power (heat) is squared in relation to voltage.

Increasing the voltage by 10% increases the power by 21%.

So yes, voltage is a big knob when it comes to heat generation, CPUs and all electrical things.