NVIDIA GF100 & Friends speculation

[neliz]

NV promised review cards 2 weeks before the launch to the press. So coming Friday we'll know a lot more.

re: A3/B1, we see quantities of A3 in April/May simply because it takes time to process the dies.
The paper launch in March is a simple indication it takes more time than they want. So any quantities we'll have in the next two months is whatever has been processed/is being processed right now. There won't be more than that until the B1 spin. NV currently has 80% of TSMC's 40nm production.

 

[Alexko]

Or you can play it like amd and put 2 different teams to fix the same problem for both the dies, sharing the work.

I guess that might work, but would require more resources. And obviously, if B1 doesn't cut it for GF100, then GF104 (since that seems to be its name) would require an additional spin as well, so there might be wasted wafers.

I'm just not sure NVIDIA can spare so many people to work on that with Fermi 2 probably on the way. Then again, Juniper is a big problem right now for them since they don't have anything to compete against it, so it might well be worth the extra effort.

 

[trinibwoy]

There won't be more than that until the B1 spin.

That's fine. I only need one

NV currently has 80% of TSMC's 40nm production.

Hmmm, how did you come by that particular number? GT2xx will be taking their fair share but I would be surprised if Fermi is monopolizing the line given the ongoing talk of poor yields. Why hold up 80% of your production capacity for a chip that isn't coming off the line in working order?

 

[Razor1]

So...any word or speculation regarding the idle power consumption of GF100 based cards?

GT200 held it's ground regarding idle power consumption, although it was a big chip. Is there a reason we shouldn't expect the same with GF100 or the opposite?

very low, it should hold its ground just fine

 

[neliz]

Hmmm, how did you come by that particular number?

quoting someone (no, not charlie)

 

[Jaaanosik]

NV promised review cards 2 weeks before the launch to the press. So coming Friday we'll know a lot more.

re: A3/B1, we see quantities of A3 in April/May simply because it takes time to process the dies.
The paper launch in March is a simple indication it takes more time than they want. So any quantities we'll have in the next two months is whatever has been processed/is being processed right now. There won't be more than that until the B1 spin. NV currently has 80% of TSMC's 40nm production.

Was there a B1 tape out reported by anybody? When it happened?

 

[rpg.314]

I guess that might work, but would require more resources. And obviously, if B1 doesn't cut it for GF100, then GF104 (since that seems to be its name) would require an additional spin as well, so there might be wasted wafers.

I'm just not sure NVIDIA can spare so many people to work on that with Fermi 2 probably on the way. Then again, Juniper is a big problem right now for them since they don't have anything to compete against it, so it might well be worth the extra effort.

Fermi2 needs to deliver, no doubt. You can survive 1 fuckup, but not 2 consecutive ones. More engineering resources have probably been shifted there. Considering that B1 is set for June/July, I think it is not too far from completion now.

 

[trinibwoy]

Fermi 2? Why does anybody expect to see that before Q3/4 next year? At best you can hope for a G92 like 28nm shrink of Fermi 1 in Q2.

 

[rpg.314]

At best you can hope for a G92 like 28nm shrink of Fermi 1 in Q2.

That will make it 4-6 months later than NI.

 

[MDolenc]

Well... Tesla 2 didn't come out a few months after Tesla did it now? Of course if you're talking about chips faster then GF100 that's a whole other story.

 

[Alexko]

Fermi 2? Why does anybody expect to see that before Q3/4 next year? At best you can hope for a G92 like 28nm shrink of Fermi 1 in Q2.

Well, Fermi was scheduled for Q3/Q4' 2009, so Fermi 2 probably was (maybe still is) scheduled for Q3/Q4 2010...

 

[rpg.314]

NV hasn't delivered any >450mm2 chip within a year of it's successor in a long time now.

 

[Sontin]

NV hasn't delivered any >450mm2 chip within a year of it's successor in a long time now.

Where is the difference to AMD?

 

[ShaidarHaran]

Where is the difference to AMD?

AMD doesn't design chips that large. NV does.

 

[Silus]

That will make it 4-6 months later than NI.

Which you are assuming to be on time...
ATI hasn't delivered a new architecture, without any kind of problems for quite some time now, plus we still have to see what Fermi 1 does, with its supposed target clocks, which almost everyone here is assuming it won't reach before B1 (and some even doubting that B1 will be enough...).

If NI delivers the same boost in performance that RV870 provided over RV770 (R600 over R580 didn't even provide that and was a totally new architecture), then Fermi with its target clocks or beyond them, may prove to be on par if slightly slower than NI, in performance. Of course this is assuming that the GTX 480 is indeed 20-30% faster than the HD 5870, as was mentioned a while back, with its "low" clocks and that with its target clocks, it can get a 50% boost over RV870.

 

[mczak]

NV currently has 80% of TSMC's 40nm production.

Wow there. For a chip that only even rarely works? Or how much of that is gt2xx?
They reserved all that capacity just so AMD can't produce enough chips or what .

 

[Dave Baumann]

The BS meter is waggling wildly at the moment.

 

[rpg.314]

Wow there. For a chip that only even rarely works? Or how much of that is gt2xx?
They reserved all that capacity just so AMD can't produce enough chips or what .

My guess would be mobile chips.

 

[rpg.314]

Which you are assuming to be on time...
ATI hasn't delivered a new architecture, without any kind of problems for quite some time now, plus we still have to see what Fermi 1 does, with its supposed target clocks, which almost everyone here is assuming it won't reach before B1 (and some even doubting that B1 will be enough...).

If NI delivers the same boost in performance that RV870 provided over RV770 (R600 over R580 didn't even provide that and was a totally new architecture), then Fermi with its target clocks or beyond them, may prove to be on par if slightly slower than NI, in performance. Of course this is assuming that the GTX 480 is indeed 20-30% faster than the HD 5870, as was mentioned a while back, with its "low" clocks and that with its target clocks, it can get a 50% boost over RV870.

1) At this point, I have more confidence in AMD's ability to execute than nv's.

2) A shrink should help achieve higher clocks.

3) A new architecture should (aka hopefully) alleviate some/many bottlenecks and inefficiencies.

4) AFAIK, rv770 was on time. And it certainly created more problems for the competition than it's maker. :smile:

 

[trinibwoy]

Well, Fermi was scheduled for Q3/Q4' 2009, so Fermi 2 probably was (maybe still is) scheduled for Q3/Q4 2010...

Q3/Q4 2010 on what process? TSMC's 28nm risk production starts in Q3 2010. http://danielnenni.com/2010/01/31/tsmc-28nm-design-advisory/