That would certainly make sense from everything I'm hearing. (Although DT isn't all that reliable)
What will replace NV50?
- Thread starter bbot
- Start date
That would certainly make sense from everything I'm hearing. (Although DT isn't all that reliable)
DegustatoR
Legend
I'm not completely sure that this NV48 is the same NV48 we were expecting 4Q04...radeonic2 said:
Yeah, it would be pretty dumb to release a 16-pipe 110nm NV48 part to put up against R520 in Q205. I'd expect at least a couple of extra quads and something else a bit special to stay competitive against ATi. I'm pretty surprised we didn't see the 'original' NV48 in this quarter - even with only a modest speed increase over NV40 they would have been cheaper to produce and with better yields too. Surely it was worth doing for the cheaper GPUs and better supply? :?DegustatoR said:
DegustatoR said:
No, I would suggest that, if anything, this is what we were pinning "NV47" as - assuming any of this is the case.
Coz said:
Correct me if I'm wrong, but isn't the NV43@500MHz consuming more power than a simple 6800nonUltra?
We've got 2 quads and 500MHz in the first case, and 3 quads and 325MHz in the second (for the record power consumption hasn't risen considerably while unleashing the locked 4th quad and oc'ing to 350MHz here). I could think of doubling the amount of quads of a standard 6800 at 325MHz to nearly double it's power consumption too. Now add the non-low-k 110nm and the possibility of higher clockspeeds than a 6k8U sounds hard to achieve.
I did some experiments with the 6k8nonU and 3dmark03 a couple of days ago and the results look like this:
- 6800LE (8p/4vp@325/350MHz)
6870 points
GT1: 187.4 fps
GT2: 52.2 fps
GT3: 44.2 fps
GT4: 38.5 fps
- 6800 (12p/5vp@325/350MHz)
8343 points
GT1: 199.0 fps
GT2: 66.5 fps
GT3: 56.1 fps
GT4: 46.2 fps
- 6800 (16p/6vp@325/350MHz)
9147 points
GT1: 202.2 fps
GT2: 74.0 fps
GT3: 62.4 fps
GT4: 51.6 fps
- 6800 (16p/6vp@350/435MHz)
10135 points
GT1: 207.6 fps
GT2: 84.2 fps
GT3: 70.2 fps
GT4: 56.8 fps
- 6800LE (8p/4vp)
VS2.0: 19.7 fps
PS2.0: 84.4 fps
- 6800 (12p/5vp)
VS2.0: 23.2 fps
PS2.0: 113.5 fps
- 6800 (16p/6vp)
VS2.0: 26.1 fps
PS2.0: 134.1 fps
- 6800 (16p/6vp@350/435MHz)
VS2.0: 28.7 fps
PS2.0: 150.2 fps
Display resolution: 1280*960*32
Antialiasing: 2x
Anisotropic filtering: 4x
- 6800LE (8p/4vp)
4159 points
GT1: 139.2 fps
GT2: 28.6 fps
GT3: 23.9 fps
GT4: 24.8 fps
- 6800 (12p/5vp)
5161 points
GT1: 154.8 fps
GT2: 35.9 fps
GT3: 30.8 fps
GT4: 32.4 fps
- 6800 (16p/6vp)
5727 points
GT1: 161.2 fps
GT2: 39.8 fps
GT3: 34.5 fps
GT4: 37.6 fps
- 6800 (16p/6vp@350/435MHz)
6432 points
GT1: 171.1 fps
GT2: 46.4 fps
GT3: 39.5 fps
GT4: 41.5 fps
More here:
http://www.mitrax.de/?cont=artikel&aid=24&page=15
Now of course are there milestones with a simple 6800 (such as half the framebuffer of larger models, less available bandwidth etc.), yet the resolutions used aren't high enough to signify bigger differences.
You may be right but don't forget that in power consumption tests the X800XL (400MHz 110nm 16-pipe) only consumed about 22 watts more than the X700Pro (420MHz 110nm 8-pipe). So all those extra transistors didn't increase power consumption by a huge amount in that case. Besides, higher power consumption doesn't mean lower clockspeeds (unless heat becomes a big problem or power supply to the gfx card reaches a limit). I think NV can manage to squeeze a few more MHz out of 110nm than ATi did. Looking at the mainstream 110nm parts, the X700 stuff barely overclocked over 500MHz but 6600GT regularly reached the 560-590MHz range. Yeah, yeah, I know we're talking about different architectures etc. here but I'm just trying to say that things may not be as bleak as you say they are.Ailuros said:
Transistor count for 4 NV4x quads and 400MHz max is at 222M; adding another 2 quads breaks any so far analogy and doesn't make it predictable, especially compared to a 2 quad NV43@500MHz.
Needless to say that it such a case scenario could indicate a quite bandwidth limited scenario (just as some indications so far look like for first R520 boards); that's what I was trying to tell with the former benchmark results. It'll boost significantly pixel shader or AF performance for instance, overall though not as much as some would expect.
Needless to say that it such a case scenario could indicate a quite bandwidth limited scenario (just as some indications so far look like for first R520 boards); that's what I was trying to tell with the former benchmark results. It'll boost significantly pixel shader or AF performance for instance, overall though not as much as some would expect.
Ailuros said:
I think with the kind of per-clock gains we're going to see in R520, such a part would have to have to hit ~600MHz to stay competetive. That's asking alot from a tweaked NV40.
Of course, it's hard to comment without knowing more about the respective yield characteristics for a given die size and process. Somehow 24 pipes@400MHz on 0.11u seems more attainable than 16pipes@600MHz on 0.13u.
True. I think the key for that riddle (and probably the primary reason for such a decision) lies within availability. Either scenario points at high power consumption and bandwidth restrictions (until faster ram becomes available).
How's the situation with 90nm at IBM and does TSMC have spare space for low-k 90nm?
DegustatoR
Legend
Well, i was expecting something like RX850, not even a shrink to 110nm but just the same NV40 with native PCIE support and some bugfixes (video processor being the biggest of them).radeonic2 said:So what was the NV48 we were expecting, and what is it likely to be now?
Sorry if this is dumb question
Now it's likely to be something bigger, something like 24 pixel pipes and 8 vertex pipes and 110 or even 90nm.
So is this supposed to come out before the R520?DegustatoR said:Well, i was expecting something like RX850, not even a shrink to 110nm but just the same NV40 with native PCIE support and some bugfixes (video processor being the biggest of them).radeonic2 said:So what was the NV48 we were expecting, and what is it likely to be now?
Sorry if this is dumb question
Now it's likely to be something bigger, something like 24 pixel pipes and 8 vertex pipes and 110 or even 90nm.
NV48 on 0.11 sounds reasonable to me. If it did indeed contain 24 pixel 'pipes', it wouldn't need a particularly high clock speed. Even if clock speed went down to say, 350MHz, it would still have a third more pixel power than the standard NV40 Ultra.
If they could release this before R520 and it had similar performance they could do very well from it. Using the tried and tested 0.11 process ought to make it easier for them to implement the chip than using the potentially risky 0.09 as it is rumoured ATI is aiming for with R520.
It's just like R300 vs NV30 except in reverse!
The question is, will ATI be able to to succeed in advancing to the smaller process as NV did in the past on a number of occasions, or with they get bitten in the arse as NV did with 0.13 low-K?
If they could release this before R520 and it had similar performance they could do very well from it. Using the tried and tested 0.11 process ought to make it easier for them to implement the chip than using the potentially risky 0.09 as it is rumoured ATI is aiming for with R520.
It's just like R300 vs NV30 except in reverse!
The question is, will ATI be able to to succeed in advancing to the smaller process as NV did in the past on a number of occasions, or with they get bitten in the arse as NV did with 0.13 low-K?
Maybe memory speed is a part of that, too? The 6800 packs only 350MHz DDR, whereas the 6600G uses relatively fast 500MHz DDR3.Ailuros said:
The GT's 500MHz core may also require a higher vcore than the 6800's relatively low 325MHz. And maybe its fuller media processor takes some more power?
Similar threads
