NV50 specs?

[kenneth9265_3]

I know it may be way too early to really talk about it, but it looking more and more likely that the NV40 is going to have monsterous-like performance...If the NV40 does turns out to have 16x1/32x0 performance, where how will this affect the NV50, a whole next gen GPU?
What do you think would be bigger jump in tech/performace, NV30-NV40 or NV40-NV50?

 

[StealthHawk]

I know it may be way too early to really talk about it

You're right.

And I say NV50 just because I think DX Next is a bigger shift than the one from DX9.0 -> DX9.0c

 

[kenneth9265_3]

I know DX10/Longhorn is going be a dramatic impact for Windows, but is it going to bring anything more to gaming as DX9 was for gaming? Other then the realistic graphics of HL2, DX9 wasn't very "revolutionary" to me. Is DX10 going to be any different? :?

 

[KimB]

Well, considering the NV30 was supposedly plagued with many fabrication problems, was possibly released as a much weaker chip than was originally intended, and was nVidia's worst showing to date as it compared against the competition, I would tend to think that nVidia will have a much easier time massively increasing performance (particularly in FP shaders) than ATI, and will definitely have an easier time massively increasing performance this time than next (provided the NV40 isn't a pile).

 

[anaqer]

Wow, almost sixty percent for the NV30-NV40 jump... there sure seem to be a lot of believers 'round here.

 

[ninelven]

Well, biggest jump in technology is probably going to be R420 to R500 IMHO, though one could make a case for NV20 to NV30 I suppose. Performance wise I'd go with NV40 to NV50.

 

[KimB]

Wow, almost sixty percent for the NV30-NV40 jump... there sure seem to be a lot of believers 'round here.

It shouldn't require a lot of believing. Just noticing that the NV30 is a much lower-performing design than previous chips put out by the manufacturer (in relative terms). At least as far as FP performance is concerned, the NV30 has far, far more to go in performance than the R300 (particularly since it has around half of the FP power of the R300, having most of its processing dedicated to integer, and that's not even taking into account the FP register usage issues).

Not to mention the NV30 has a 128-bit memory bus while the R300 has a 256-bit memory bus. That alone should make it much easier for nVidia to raise the performance of the NV40 over the NV30 quite a bit more than ATI.

So, all that voting that the NV30 vs. NV40 will be a bigger performance jump than R300 vs. R420 requires is belief that the NV40 will be competitive with the R420 in all aspects of performance. Given nVidia's past of directly attacking previous architecture's main weakspots, I think this is likely.

 

[Ailuros]

That poll is quite one-sided. Where are the R200-R300 and R420-R500 options?

 

[anaqer]

It shouldn't require a lot of believing.

Oh it does alright.

Putting the NV30-NV40 transition ( in terms of performance / feature difference, of course ) before the NV40-NV50, of which we know nothing, zinch, nada speaks volumes of the near-blind faith people seem to have in the NV40. How can you be so sure this will be a bigger jump than one that isn't bound to happen for... oh dunno, like, 18 months maybe?

Noticed how I wasn't even casually mentioning the R420 OR the R500?


EDIT : fixed typo and inserted "feature" ( DXNext, anyone? )

 

[kenneth9265_3]

sorry, I could post was 4.

 

[KimB]

It shouldn't require a lot of believing.

Oh it does alright.

Putting the NV30-NV40 transition ( in terms of performance / feature difference, of course ) before the NV40-NV50, of which we know nothing, zinch, nada speaks volumes of the near-blind faith people seem to have in the NV40.

No, not at all. It just requires a realistic view of what is possible, and a believing that the NV30 was plagued with many problems, and thus wasn't as high-performing as the original design should have been (this should be easy to believe if one also believes the statements that the original NV30 design was meant to be manufactured on a low-k process).

If the NV40 is "all it can be," regardless of whether or not it beats out the R420, it will be much faster than the NV30. Given this it is unlikely to see as much of a leap with the NV50.

Then there's the other piece of general evidence: current high-end chips are hot. Very, very hot. As a consequence they also draw lots of power. This just can't continue. Chips will have to make sacrifices somewhere to bring the heat down.

And of course, there's the problem of physics. Put simply, we are coming all too close to the absolute limits of silicon transistor technology, and as such further reductions in size will become slower and slower in the near future. Once we find some radically new technology to rely upon, that could all change, and movement to other semiconductors (e.g. diamond) could speed things up again for a short time, but right now, silicon semiconductor advancement is slowing.

Now, what I don't understand is why people would vote that there will be an even larget jump from NV40 to NV50, because, as you stated, there is no information about the NV50, and given the general arguments laid out above, we would need to have specific information to the tune of, "The NV50 will use X technology to dramatically surpass the NV40," for there to be any reason the NV50 will outpeform the NV40 by more than the NV40 will the NV30.

 

[kenneth9265_3]

Maybe they voted that way because of the belief that DX10 will be a factor?

 

[KimB]

Maybe they voted that way because of the belief that DX10 will be a factor?

A massive change in architecture is more likely to reduce the performance differential, at least until software comes out that makes adequate use of the new architecture.

 

[anaqer]

You have a valid point about the limits of the current technology, but I think these limits can be circumvented ( or at least pushed back ) with a somewhat different design philosophy and/or new foundry processes ( I believe Intel has something going on that's due 2005 - 2006...? ).

And yes, I do believe DXNext to be of a wee bit more significant than going from shader model 2+ to shader model 3.


EDIT : uhmm... I may have misunderstood the poll. Are we talking about performance in prev-gen applications alone or performance AND added new features?

 

[ChrisRay]

It shouldn't require a lot of believing.

Oh it does alright.

Putting the NV30-NV40 transition ( in terms of performance / feature difference, of course ) before the NV40-NV50, of which we know nothing, zinch, nada speaks volumes of the near-blind faith people seem to have in the NV40.

No, not at all. It just requires a realistic view of what is possible, and a believing that the NV30 was plagued with many problems, and thus wasn't as high-performing as the original design should have been (this should be easy to believe if one also believes the statements that the original NV30 design was meant to be manufactured on a low-k process).

If the NV40 is "all it can be," regardless of whether or not it beats out the R420, it will be much faster than the NV30. Given this it is unlikely to see as much of a leap with the NV50.

Then there's the other piece of general evidence: current high-end chips are hot. Very, very hot. As a consequence they also draw lots of power. This just can't continue. Chips will have to make sacrifices somewhere to bring the heat down.

And of course, there's the problem of physics. Put simply, we are coming all too close to the absolute limits of silicon transistor technology, and as such further reductions in size will become slower and slower in the near future. Once we find some radically new technology to rely upon, that could all change, and movement to other semiconductors (e.g. diamond) could speed things up again for a short time, but right now, silicon semiconductor advancement is slowing.

Now, what I don't understand is why people would vote that there will be an even larget jump from NV40 to NV50, because, as you stated, there is no information about the NV50, and given the general arguments laid out above, we would need to have specific information to the tune of, "The NV50 will use X technology to dramatically surpass the NV40," for there to be any reason the NV50 will outpeform the NV40 by more than the NV40 will the NV30.

This is an issue thats really starting to bug me. Both ati AND nvidia run way too hot on their high end cards.

High End Cards + High End PC = Non Convention Heater During Winter

 

[malcolm]

I think nv40 to nv50 because the nv50 will probably be on 90nm right?
This would probably mean twice the amount of transistors so it seems pretty obvious it should be a lot faster.
If the memory improvement isnt anough to keep up maybe this is a good time to implement analytical aa?

 

[BRiT]

My choice would be R420 to R500 as way of technology jump.

But of the choices presented, I think NV40 to NV50 will have a largest performance delta on the Nvidia side. Not because I think the NV50 will be amazing, but because I think the NV40 won't perform that well.

 

[Megadrive1988]

NV50 guesstimate
*512 Bit Bus
*32 pixel pipelines (32x1, 64x0)
*16~32 MB embedded memory
*512 MB external memory
*VS/PS 4.0+
*1~2 billion verts/sec peak

 

[KimB]

This is an issue thats really starting to bug me. Both ati AND nvidia run way too hot on their high end cards.

High End Cards + High End PC = Non Convention Heater During Winter

It's kind of unavoidable when you want the best performance possible. High-end cards are always going to be hot from here on out. The same goes for CPUs.

 

[KimB]

NV50 guesstimate
*512 Bit Bus
*32 pixel pipelines (32x1, 64x0)
*16~32 MB embedded memory
*512 MB external memory
*VS/PS 4.0+
*1~2 billion verts/sec peak

512-bit bus highly unlikely. 256-bit is already pushing the limits, and 512-bit will just be more expensive.