As far as I'm concerned, AMD won't release a new series on 20nm at the end of this year. They already went through this with TSMC once before when Northern Islands was cancelled for 32nm. We know they'll be releasing *something* though so the smart money would be on a similar situation to Northern Islands, that is the new series on the same node.
If by some miracle TSMC actually manage a 2-year cadence this time around, it's feasible that Nvidia will be ready with Maxwell on 20nm before the middle of 2014. For me it's a lot more likely that Nvidia learned from AMD's mistake and Maxwell will be seen on 28nm first as well.
If by some miracle TSMC actually manage a 2-year cadence this time around, it's feasible that Nvidia will be ready with Maxwell on 20nm before the middle of 2014. For me it's a lot more likely that Nvidia learned from AMD's mistake and Maxwell will be seen on 28nm first as well.
I personally think that both Nvdia and AMD will release their next gen cards on 20nm and not on 28nm. JHH pretty much confirmed this for Nvidia back in April when Nvidia hosted their annual investors day "Maxwell will crush Kepler on efficiency" and that simply cannot happen if Maxwell is a 28nm part. 20nm is on track for an early ramp. TSMC said there will be more 20nm volume this year and next year than there was with 28nm in 2011 and 2012.
I think the biggest issue that will be holding back next gen releases is the finalization and mass production of GDDR6 memory. Performance of today's cards are more or less at bandwidth limits, especially with Kepler. Unless either camp plans on moving to ridiculously large bus sizes (which will inefficiently take up more die space in controllers and can be potentially problematic) then both Nvidia and AMD will be slave to GDDR6 being viable.
If GDDR6 isn't in mass production by the end of this year or early 2014, then I think we are more likely to see die shrinks before new architectures.
I think the biggest issue that will be holding back next gen releases is the finalization and mass production of GDDR6 memory. Performance of today's cards are more or less at bandwidth limits, especially with Kepler. Unless either camp plans on moving to ridiculously large bus sizes (which will inefficiently take up more die space in controllers and can be potentially problematic) then both Nvidia and AMD will be slave to GDDR6 being viable.
If GDDR6 isn't in mass production by the end of this year or early 2014, then I think we are more likely to see die shrinks before new architectures.
iMacmatician
Regular
I wonder how the situation is compared to 2006-2008 and GDDR3. NVIDIA went to 512-bit with the GT200, so even if it's difficult, maybe they'd go to 512-bit on, say, "big" Maxwell if they were backed into a corner on memory bandwidth for it. (Knights Corner already has a 512-bit bus with 5.5 Gbps GDDR5.)
What would be the costs of 20 nm chips compared to 28 nm? IIRC AMD moved the in-development 32 nm NI chips, except for the high-end part, to 40 nm even before 32 nm was canceled, since the chips would be cheaper on 40 nm. Also, if a 28 nm chip was shrunk to 20 nm then it could have similar or greater FLOPS but not necessarily (much) more bandwidth (at least for those that are at or near 6 Gbps already). It might even have less bandwidth if the shrink means that fewer memory interfaces can fit on it. So shrinks may not give much more performance (although that may not be the goal of a particular shrink) but should have a lot higher power efficiency.
More memory, far better compute, and judging from titan, smoother, more consistent frame rates. I know that last one may not necessarily scale down, but comparing 680 to Titan when they are at their closest, I think it would still be a factor.
The problem though is that both companies would need memory bus increases across all their products, not just the flagship gpu's. And then we'd be looking at mid-range dies with 384-bit buses because 320-bit is an odd number for memory configurations (assuming 3gb is the norm for vram), unless Nvidia wants to mix-max memory configs for most of their product line which I would not think is ideal.
It's just me musing out loud. I don't think maxwell GM104 will be released until May or June of 2014, so I think nvidia has plenty of time. I also don't see AMD releasing their 20n, flagship until March or so. AMD will probably have a leg up on gddr6 controllers since they have a direct role in its development.
Last edited by a moderator:
According to the extremetech article, Nvidia believes that it'll be Q1 2015 before 20nm becomes worth it in cost terms.
http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless
Even if TSMC has pulled 20nm in by 6 months, you'd still be looking at Q3 2014. If AMD releases this year on 20nm the cost of the cards are going to be astronomical surely?
http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless
Even if TSMC has pulled 20nm in by 6 months, you'd still be looking at Q3 2014. If AMD releases this year on 20nm the cost of the cards are going to be astronomical surely?
I haven't seen that much of a difference between GK104 and GK110 in compute, except of course in DP, and when Hyper-Q or dynamic parallelism are used, but I don't expect that to matter much in games.
If you're using CUDA for GPGPU, of course, that's a different story (assuming the GTX 780's DP is unlocked, as Titan's is, which is not guaranteed).
You shouldn't forget the 50% higher ROP count, the 50% higher TMU count and also not the higher geometry throughput.
Those slides can't be relied on, they are based on out of date information.
Also i thought Nvidia released these slides to put pressure on TSMC?
how can an hypotetic gtx880 goes better than Titan, or at least gtx780, if the chip is on 28nm?
A1xLLcqAgt0qc2RyMz0y
Veteran
According to the extremetech article, Nvidia believes that it'll be Q1 2015 before 20nm becomes worth it in cost terms.
http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless
Even if TSMC has pulled 20nm in by 6 months, you'd still be looking at Q3 2014. If AMD releases this year on 20nm the cost of the cards are going to be astronomical surely?
20nm will be worth it from the start for Nvidia especially for the high-end parts where the price is sky high. Even from the charts shown parity in price with 28nm will occur in 2013Q3 and with TSMC pulling in 20nm production 3-6 months that means the chart should slide forward 3-6 months.
I expect the split in Kepler between the GK110 and the GK104 to again happen with Maxwell so the higher initial cost of 20nm would be somewhat muted.
Also if any of the rumored 28nm Maxwell's come to be it would be in the low end of the stack and be replaced with 20nm parts when price parity occurs.
Well TSMC claimed they'd moved forward 2 months, but recent talk says they won't be entering volume production on 20nm until 2014 - http://www.digitimes.com/news/a20130510PD204.html?mod=3&q=TSMC
They are also still expanding 28nm all year. Looking at the numbers I would be surprised if both Nvidia and AMD felt it was a lot more economically viable to wait it out with another year on 28nm.
Nvidia might be going 20nm in mid 2014 but I cannot see AMD going 20nm in Q4 2013. Not unless the really plan to surprise us with something extraordinary like a gigantic GPU that they can command $1K for. I haven't read anything about 20nm tapeouts either.
They are also still expanding 28nm all year. Looking at the numbers I would be surprised if both Nvidia and AMD felt it was a lot more economically viable to wait it out with another year on 28nm.
Nvidia might be going 20nm in mid 2014 but I cannot see AMD going 20nm in Q4 2013. Not unless the really plan to surprise us with something extraordinary like a gigantic GPU that they can command $1K for. I haven't read anything about 20nm tapeouts either.
Last edited by a moderator:
It doesn't have to be faster than Titan, that's why Titan got the name. Also, Nvidia don't have to be releasing the GTX 880 first - there's no reason why they wouldn't be planning their original plan this time around and releasing the 860 Ti or 870 first.
no reasons until Amd will release her 20nm based vgas, and what about if amd will release those vgas before nvidia, as it was for 7970...?
Nvidia is used to fighting AMD's cards with previous-gen offerings for a while. The 560 Ti was the competition for the 7850 for over 6 months, with various other 560 cards competing vs the 7770 for even longer until they finally released the Kepler competition some 9 months or so later.
If we assume AMD does release a 20nm gpu, say slightly smaller than Tahiti and 15-20% faster than Titan, Nvidia wouldn't really be in much worse shape than they were for the majority of last year. They could also release the full shader Titan and probably get near to a draw.
The only risk they'd be taking would be if AMD had a mammoth gpu and TSMC's process was spectacular from day one. Even one of those seems unlikely but both would be pretty bizarre.
UniversalTruth
Veteran
But it will be. There is no need always to look for something unexpected and out of any logic.
As to how it can be faster and still on 28 nm- simply by using a better architecture.
Doesn't make sense again. This name is just randomly put there just to throw the particular card out of any series. To justify the way NV screws customers by not providing top performance for a reasonable price...
Silent_Buddha
Legend
Eh? The chart doesn't show the predicted 20 nm cost line (per transistor) intersecting with the 28 nm cost line until 4Q2014. That's quite a ways away.
Regards,
SB
