NVIDIA GF100 & Friends speculation

Another shipment is scheduled to go out before the end of the year, bigger than last time as well.

Just got an email from the guy, he says since Nvidia are stuck on 40nm for the next year or so, they are planning a 640SP card as well. Apparently it will use 40SP SMs, with 4 SMs per GPC and a 4 GPC cluster, they are working to pack the chip more densely so they can add the SPs without making it any bigger.

The dude is very credible, he has given me a lot of information before about other stuff from within the industry. I just want to know if a card like that would be possible, it seems like it would be to me.
 
Or maybe gamers got tired of waiting for AMD's "now known lower performance" 6970 and finally bought the GTX580 thus depleting stocks.

Looks like nVidia has been minting money on the GTX580 since early November.


Nvidia is looking good, while AMD is not. I don't know, but that's the first impression here. Me also is dissapppointed of that so low number of SPs for R6970. Only 1536 SPs? :oops: I hope they are building something more impressive, because NVidia might have an answer even for Antilles. :oops:
 
Another shipment is scheduled to go out before the end of the year, bigger than last time as well.

Just got an email from the guy, he says since Nvidia are stuck on 40nm for the next year or so, they are planning a 640SP card as well. Apparently it will use 40SP SMs, with 4 SMs per GPC and a 4 GPC cluster, they are working to pack the chip more densely so they can add the SPs without making it any bigger.

The dude is very credible, he has given me a lot of information before about other stuff from within the industry. I just want to know if a card like that would be possible, it seems like it would be to me.

I think the guy is either misinformed or pulling your leg:

  • A third type of SM with 40SPs, really?
  • It just wouldn't fit within 600mm² on TSMC's 40nm process;
  • JHH was pretty clear: one "mid-life kicker" [GF110] and then Kepler in 28nm.
 
Last edited by a moderator:
Just got an email from the guy, he says since Nvidia are stuck on 40nm for the next year or so, they are planning a 640SP card as well. Apparently it will use 40SP SMs, with 4 SMs per GPC and a 4 GPC cluster, they are working to pack the chip more densely so they can add the SPs without making it any bigger.
I call BS on that. Now, GF100/GF110 and the rest of GF1xx are already quite dissimilar due to the different dispatch (needed by the 32SP vs. 48SP thing), but both feature 16-wide alus. Changing that to 8-wide so 40SP are possible would be a step back towards earlier designs, which imho makes no sense.
Unless you're talking completely new chip, but in that case it doesn't really make sense to talk about "packing more densely".
 
Or maybe gamers got tired of waiting for AMD's "now known lower performance" 6970 and finally bought the GTX580 thus depleting stocks.

Looks like nVidia has been minting money on the GTX580 since early November.

We got word in Finland sometime last month or very early this month that next shipment of GTX580's is expected to arrive in any nordic country sometime in January
(Only from one big manufacturer though, but still gives some view on the situation)
 
I think the guy is either misinformed or pulling your leg:

  • A third type of SM with 40SPs, really?
  • It just wouldn't fit within 600mm² on TSMC's 40nm process;
  • JHH was pretty clear: one "mid-life kicker" (GF110) and then Kepler in 28nm.

Yeah, I asked those questions too.

1. 48SPs doesn't work for 640SPs as it gives an odd number of SMs, and using 32SPs in an SM would give 20 SMs total which is too difficult to split properly.
2. Nvidia are working on making the chip more densely packed so that they don't use more silicon than they currently do, ATi's chips are, for example, 20-25% more dense than Nvidia's on the same process node.
3. That was before TSMC silently delayed 28nm into 2012. Aparently both ATi and Nvidia are stuck with 40nm until at least late 2011 more likely early/mid 2012. So plans had to be changed, and he is saying this is what they have come up with.

A new chip with 640SPs should provide 20-30% more performance for the same silicon budget if Nvidia's work goes to plan.

Take a look at the chip layout of GF110, he said that Nvidia need to add one more column of SPs to each SM and it would yield a lot of other benefits as well most of which I didn't understand but something to do with the SMs being wider.
 
So I called the guy and he said 200W was for Furmark, 150W was for gaming performance. I misunderstood what he said.

He gave me the specs again to make sure I had it right:

384SP
775-820MHz
150W TDP, 200W Furmark
350-370mm^2 die
256bit, 1GB.

30 - 40% better than the GTX460 1GB according to Nvidia, but he says 20-30% is more realistic.
I think if GF110 was really pretty much GF110b, then GF114 will be rather GF104 A3 (ok I don't know the stepping of shipping GF104 but you get the idea).
775Mhz sounds quite doable - all GF1xx chips are apparently designed to reach about that clock.
The power numbers also sound credible - I would pretty much expect an increase corresponding to the increased performance (the GTX570/580 did better than that, but mostly due to better cooling and solving some leakage issues, both would not apply to GF114 based cards). Even overclocked to 810Mhz GTX 460 stay below 150W for games and below 200W for furmark, adding 10% or thereabouts for that additional SM would put them right there.
Even the performance numbers sound credible - 30-40% better looks indeed exaggerated (at 775Mhz theoretical alu improvement is 31%, so if clock is a bit higher 40% are possible too but practical numbers will stay below as usual). Would be a nice HD 6870 competitor.
 
In Asia rumors are spreading of a 875MHz 384SPs 256-Bit 185W TDP GTX 560:
http://translate.google.de/translat...ad-38537-1-1.html&sl=zh-CN&tl=en&hl=&ie=UTF-8
http://we.pcinlife.com/thread-1572757-1-1.html

Clock seems a bit high, but if they can achieve this, this should be a HD 6950 competitor.
Dunno, more sounds like some OC card to me. 875Mhz are not entirely impossible with GF1xx chips, some factory OC GTS 450 ship with that clock, but they (afaik) all have higher than reference voltage (and even the reference one has not so good power draw considering performance because the voltage is higher than with GF104 chips). If that's games TDP, and (unrestricted) furmark TDP is more like 230W (and at this point it will exceed HD 6950 power consumption), I say could be, but otherwise looks impossible to me.
 
Yeah, I asked those questions too.

1. 48SPs doesn't work for 640SPs as it gives an odd number of SMs, and using 32SPs in an SM would give 20 SMs total which is too difficult to split properly.

Sure, it's the only way the math works out, but that doesn't mean it makes sense from an engineering point of view.

2. Nvidia are working on making the chip more densely packed so that they don't use more silicon than they currently do, ATi's chips are, for example, 20-25% more dense than Nvidia's on the same process node.

They spent months redoing the layout for GF100, and what we ended up with is an even bigger chip, around 550mm² according to Charlie. AMD's GPUs are denser, but they also lack anything comparable to NVIDIA's hot clock domain, so they can pack more transistors in the same area. There's no such thing as a free lunch.

3. That was before TSMC silently delayed 28nm into 2012. Aparently both ATi and Nvidia are stuck with 40nm until at least late 2011 more likely early/mid 2012. So plans had to be changed, and he is saying this is what they have come up with.

Has that delay even been confirmed? Even if it's true, TSMC was still saying something like Q2'2011 just a few months ago. I don't think that's enough time to start working on a whole new design.

A new chip with 640SPs should provide 20-30% more performance for the same silicon budget if Nvidia's work goes to plan.

Take a look at the chip layout of GF110, he said that Nvidia need to add one more column of SPs to each SM and it would yield a lot of other benefits as well most of which I didn't understand but something to do with the SMs being wider.

That doesn't make much sense, they'd have to completely re-engineer the SMs, which would have a significant impact on the rest of the chip.

I really don't buy this.
 
Wouldn't 1.5x GF114 do the trick? 3SMs per GPC - if such SM is 25% bigger than an original one, then the whole GPC should be smaller. It doesn't have to do compute, so cut the ECC, use the transistors to improve the IMC so it can handle faster chips. Along with arch tweaks, it should end up smaller than a GF100b, right?
 
That only gives 576 SPs which is not really a big enough power increase over GF110. This guy was saying 640SPs was chosen because it was a big enough increase in power over GTX580, but Nvidia would be able to fix it so it didn't go over the 24.5mm limit that TSMC impose by increasing the density of the chip. Maybe they are going to make sacrifices to the hot clock to achieve this, I don't know. I've forwarded the post above to him maybe he will reply soon and I can explain further.

I know that he is very well connected at a number of companies. I have gotten very accurate BoM information about gaming products from him before, he was the first to alert me that Sony had broken into profit making territory well before Kaz said anything at the investor meeting and before their results were announced. He told me that Nvidia would make a loss and that Intel are struggling more than they would like to let on with 22nm which could lead to a slower roll out of 22nm products than they would like. A few weeks ago he said Intel would announce the integration of 450mm wafers into their fabs. Though I haven't heard anything about the latest Intel stuff I assume it will come out sooner or later.
 
FWIW I think 640SP would be doable with 20 SMs. Just use 5 SMs per GPC - I see no fundamental reason why this couldn't work, after all it seems it works with pretty much any number (equal or lower than 4) of SMs per GPC already. Also, nvidia COULD grow the chip a bit potentially (about 10%) I think, so with 25% more SMs but everything else mostly the same, it could fit. I don't believe they could pack it more densely easily (it is not really that much less dense than AMDs chips btw), at least not by much, even with minor clock sacrificies (and I think sacrificing hot clock would imply sacrificing base clock too).
But it would likely also have power consumption problems. A 600mm² chip, similar to GF110, but just 5 SMs per GPC instead, and (say) 10% lower clock than GTX 580 might be possible within that 300W envelope. But it would only be about 10% faster than GTX 580.
 
So I called the guy and he said 200W was for Furmark, 150W was for gaming performance. I misunderstood what he said.

He gave me the specs again to make sure I had it right:

384SP
775-820MHz
150W TDP, 200W Furmark
350-370mm^2 die
256bit, 1GB.

30 - 40% better than the GTX460 1GB according to Nvidia, but he says 20-30% is more realistic.

And Nvidia will have 2x GF114 going against Antilles?

Ok I guess not.
 
Last edited by a moderator:
Or that the shipments are really small, or were just returns/sales that fell through.
Your logic does not resemble earth logic.

The GTX580 was in stock at Newegg from launch day November 9th. Yesterday was the first day with no stock. That is about 5 weeks of constant supply.

Three EVGA models were restocked yesterday and within hours sold out. Then a Galaxy & PNY model became available then also sold out within hours.

Your logic that the EVGA models were returned and then resold as Galaxy or PNY makes no sense what so ever.

What makes sense is that there is high demand for the GTX580 before the holidays and that depleted the inventory and when some cards make it back to inventory they are immediately bought up.
 
Your logic does not resemble earth logic.

The GTX580 was in stock at Newegg from launch day November 9th. Yesterday was the first day with no stock. That is about 5 weeks of constant supply.

Three EVGA models were restocked yesterday and within hours sold out. Then a Galaxy & PNY model became available then also sold out within hours.

Your logic that the EVGA models were returned and then resold as Galaxy or PNY makes no sense what so ever.

What makes sense is that there is high demand for the GTX580 before the holidays and that depleted the inventory and when some cards make it back to inventory they are immediately bought up.

And how many cards did they get in stock? You don't have their sales statistics and you're just projecting a fluctuating stock situation as a huge supply with more massive demand. The situation could just as easily be moderate demand with a tiny supply. A few cards a day is nothing for a huge e-tailer like newegg.
 
Back
Top