Since NV keeps releasing brand new architectures every 2 years but AMD is using the older style approach where their design architectures to last 4-5 years (remember VLIW lasted from 2006 all the way to 2011!), AMD starts off good in the first 1-2 years of a new architecture but then it starts to age considering NV brings out a brand new architecture 2 years after. By the time Pascal is released in 2016, NV will have gone through 3 brand new architectures in 5 years (2012-2016), while AMD will still be on GCN, which traces its roots back to December 2011. It's a very risky strategy that AMD has implemented to rely on the same architecture and improving it over the course of 4-5 years to remain competitive.
I understand why AMD needed to create a new flagship chip to maintain their image and they can reuse it as shrunken 14nm HBM2 mid-range chip for next gen but they can't fit a 250-300W Fiji chip into laptops. What's their plan a 2048 shader Tonga XT again? They really need a 2560 or even 2816 shader card in laptops at 100-125W TDP and I have my doubts this is happening with R9 300M series.
The fact that AMD is continuing to use GCN isn't the problem. The problem is that AMD blew a large portion of its R&D budget on the assumption that 20nm would be viable, and when it turned out it wasn't, there was no backup plan.
GCN 1.2 is actually fairly competitive if you look at the top binned chips (R9 M295X). These are more powerful than the GTX 960 for about the same TDP (125W). The problem is that AMD only releases the trash silicon to AIBs, so the desktop R9 285 looks really bad. (The fact that it has 2GB of RAM instead of 4GB only makes matters worse.) And the current Tonga chips are done on TSMC; the low-leakage GloFo process designed for GPUs should improve perf/watt at least marginally.
In retrospect, AMD should have had a full GCN 1.2 lineup (at least 3 chips, replacing Hawaii, Tahiti, and Pitcairn) available in mid to late 2014, and all fabbed on the GloFo 28nm SHP process. Then they would have had competition ready for big-core Maxwell, and Fiji in mid-2015 would just be the crowning touch on top, not the pillar on which the entire GPU division's destiny rests.
I can't prove it, but I think what happened is that AMD was going to do the GCN 1.2 cards on TSMC 20nm. It failed, and AMD had to scramble to put Tonga into production on TSMC 28nm via a reverse die shrink, since it had already been promised to Apple for the Retina iMac and couldn't be delayed without a major penalty. This threw AMD's already precarious R&D budget out of whack, and very little was left for what needed to be done.
Hopefully AMD has at least scraped up enough to port the chips over to GloFo for the 300 series, or else they'll be going through the next 18 months or more with absolutely nothing worthwhile for most enthusiasts in either CPUs or GPUs.
Facebook
Twitter