Amd Volcanic Islands details

[raghu78]

Kepler was launched in March 2012. You're telling me VA is meant to compete against an architecture that's already a year and a half old? If that's the case, AMD will be in serious trouble.

GK104 is a year and half old. GK110/Titan will be 8 months old by late Oct and more importantly GTX 780 will be just 5 months old. We all know once GTX 780 released, Titan became less attractive. For all intents Nvidia could launch only GTX 780 with the kind of high volumes that a desktop part needs. remember these are not low volume Quadro parts. AMD seems to be aiming for a mid October launch for Hawaii. I expect Nvidia's yields would be better by October and they could easily launch a 2688 sp GTX 785 or even better a 2880 sp full enabled Geforce chip depending on how good Hawaii is.

At this point, after skipping over their planned GCN refresh, they need to prepare for what's coming rather than what's already been available for a long time. Think of it this way. If AMD launches a 28nm VA in October, they will likely have a 6-9 month buffer zone before NVIDIA starts talking Maxwell in earnest. In order to compete with Maxwell, they'd need a chip that's (on paper at least) TWICE as powerful as TITAN if NVIDIA's latest slides are any indication of potential performance. Since refreshes typically garner at most 20% better performance than their predecessors (see previous reviews), AMD would have to come out of the gate extremely strong or at least keep something waiting in the wings to respond to Maxwell once VA gets its refresh.
I really have to wonder which is better for AMD though; wait a little while for 20nm (6 months) and compete against Maxwell on slightly more level footing or launch something now that still uses the 28nm node and hope NVIDIA doesn't pull in their launch timeframe due to strong TSMC yields. The real question is whether VA is a GCN refresh or a brand new architecture. I'm hoping for "refresh" so they can quickly transition to a new 20nm architecture in time to compete against Kepler.

TSMC 20nm volume production is Q2 2014 with actual volume of 20nm production being just 2% of total wafer volume. read TSMC Q1 2013 earnings call. see page 17 and 18

http://www.tsmc.com/uploadfile/ir/qu...Transcript.pdf

at 20nm planar the foundries are not able to provide traditional improvements in transistor perf or power gains. The transistor perf improvement from 20nm vs 28nm is 30% while from 28nm vs 40nm is 45%. 28nm also had better power gains vs 40nm as it was the first node with high k metal gates. Thats the reason that foundries are going FINFET a year later at the same 20nm node. Yields are also very difficult at the leading edge.

http://www.tsmc.com/english/dedicate...ology/20nm.htm

"TSMC's 20nm process technology can provide 30 percent higher speed, 1.9 times the density, or 25 percent less power than its 28nm technology."

http://www.tsmc.com/english/dedicate...ology/28nm.htm

"The 28HP process supports a 45 percent speed improvement over the 40G process at the same leakage/gate. "

so AMD have taken the prudent step to go with Hawaii on 28nm in Oct 2013. They might run a pipe cleaner entry level or mid range product on 20nm first releasing it by late Q2 2014 and then come out with their 20nm big chip by Sep 2014 when yields are better and 20 nm volume is significant. remember Nvidia launched GTX 680 around 2.5 months after HD 7970 and they did very well. Also AMD has clearly stated that GCN is the foundation for their GPU product lines/families for a long time to come. the R600 architecture lasted 4.5 years starting from HD 2900 XT in mid 2007 to HD 6970 till end 2011. improvements were made all along adding DX11 support etc and going from VLIW 5 to VLIW4 but they were based on the foundations of R600. GCN will be optimized and improved but will last atleast till 2016 - 2017. GCN has proved its efficiency as Pitcairn (212 sq mm) is faster than GK106(221 sq mm). Bonaire is also very efficient. Hawaii will reinstate that at the high end.

Also I believe the days of 500 - 550 sq mm on a immature leading edge process on launch are gone. Nvidia learnt that with Fermi and Kepler proved that doing massive die (500+ sq mm) on a new process is not feasible. so doubling Titan performance will not happen with GM104. given 20nm's less than traditional node gains I think even beating a fully enabled GK110 by 25% would be a massive achievement. It would mean that Maxwell is a much more efficient architecture than Kepler. GM110 on TSMC 16 FINFET in H2 2015 can double GK110 performance. so AMD's goal with their first 20nm chip should be to compete favourably with GM104. I think that should be possible. Also if AMD moves aggressively to GDDR6 while Nvidia doesn't with their first 20nm flagship GPUs then we could have a situation where AMD leapfrogs Nvidia.

AMD needs Hawaii to match Titan or slightly exceed it at around USD 600. that should be enough to force Nvidia to cut prices on GTX 780 and launch newer SKUs like GTX 785.