GTX 460 Reviews

Yep, I don't think GF104 has a chance at holding the crown. Best case is a GF100 refresh that shakes things up. Even if GF100 turned out better than it did I have to believe Nvidia would have had a 40nm refresh in the pipeline anyway. Winter 2011 is a long way away.

 

Trouble is they have very little margin for improvement in terms of die area, and none at all in terms of power.

Basically, the only thing they can do is improve efficiency. And they would have to improve it a lot to be competitive with Southern Islands, so maybe it would make more sense to release a full, overclocked GF104 to take care of replacing the GTX 470 and, to some extent, 480, and then focus on the 28nm replacement.

Why winter 2011?

 

Agreed, but they don't have a die area problem. Like you said they have an efficiency problem - if it was faster at games than it is now nobody would care about die size or power consumption. So the question is whether a variation of GF100 in the same transistor budget could produce materially higher performance and/or clocks.

Not ideal but yeah that could work in conjunction with a dual-GF104 card to make some sort of claim to the high-end.

Given that it's TSMC and Nvidia I consider that an optimistic estimate.

 

Its efficiency can be better - fully enabled GF100 at higher frequencies could be at least 20% faster than GTX480. But they couldn't release such product, because of yields and power consumption. ATi adressed these issues by doubling vias in R8xx GPUs. According to ATi, it increased die-size. nVidia could probably double vias in GF100, too, but how big would the result be? Maybe they have kind of indirect die area problem...

 

edit: Already answered

 

And most of all, can they do it without increasing power consumption at all? Otherwise they would break PCI-E specifications. Seems like a pretty tall order.

Yeah, I guess…

 

Well they would have to take the scalpel to it of course. I'm probably way off and whatever they do will probably be very different but I think a 12 SM GF104 variant would work well though there'll be a hit to geometry throughput. A 256-bit bus would help it slim down too but that assumes Nvidia could get going with higher memory speeds and that doesn't seem to be their strong suit at the moment.

 

A 3 GPC, 384-bit GF102 would be pretty interesting, with 12 SMs. Make the SMs like those in GF104?

 

Wouldn't that be severely lacking in the fillrate department (texture and various pixel)?

 

I don't see how, particularly if the SMs are the same as GF104's. The ROP count is the same as GF100.

 

Texture no, it would have 96 TMUs which should be fine. Fillrate is a potential problem if the path between the SM's and ROPs is left at 64 bits wide. That's the only thing preventing me from speculating on a 8 SM replacement for GF100 with each SM having 64 ALUs. Geometry performance would suffer but that's being artificially handicapped on consumer parts anyway according to Damien. Fillrate would be a bigger problem with just 8 SM's.

According to Anand a GF104 SM is 25% larger than a GF100 SM after adding a third shader block and doubling up dispatchers, TMUs and SFUs. Which means that third shader block is at most .25/1.25=20% of a GF104 SM. Adding a fourth shader block would take the new SM to 1.2*1.25=1.5x the size of a GF100 SM. 8 of the new SMs would take up 8*1.5/16=75% of the die area that 16 GF100 SM's use and that's estimating on the high side. Shader performance would take a hit due to the loss of efficiency with super-scalar issue but the benefits to power consumption and hence clocks could compensate somewhat. It would be another case of G80->G92, smaller die, same performance and a dual-GPU card becomes a practical consideration.

Oh, and of course for that to work they would need to beef up their register file bandwidth to feed all those ALUs.

 

NV is using double vias on GF100. I don't know to what extent (what areas) though.

 

Are you sure about that?

 

I am sure ...

http://forums.anandtech.com/showpost.php?p=30283881&postcount=219

 

Thanks! I wonder why GF100's yields are so terrible then…

 

I guess you can not cheat the size. AMD had issues even with double vias at the beginning.

 

Yep, AMD felt the squeeze on 40nm too, hence the 5850, 5830 and the 5750 having disabled units. Last generation only the 4830 had units disabled and that came much later in the product cycle. The 4850 and 4650 just had lower clocks.

 

This is an incomplete conclusion. Sure, the process changed, but so too did the architecture.

 

I meant in context of the GF100 yield discussion. Can't really argue that AMD's architecture saw much change when you look at what Nvidia did so if that's a reason/excuse it's less applicable to Evergreen.