Originally Posted by Mahigan
I didn't. I wasn't even interested in Fiji. Once Fury-X was launched I looked at the Front End and wasn't at all interested in the GPU. I saw Fiji as being a good compute part... for accelerating OpenCL amongst other things. I saw DX11 as limiting what GCN could do... and I was right.
Now we're heading toward 16nm with DX12. I think it goes without saying that AMD will improve the front end of its next gen architecture. When I couple this with the gains we're witnessing under DX12, with GCN, then I can't help but get the idea that Greenland is going to surprise a heck of a lot of people (remember what is limiting Fiji's performance and I think you'll sort of see where I'm going with this). Fiji was simply a maxed out Hawaii/Grenada whereas Greenland is a new ISA. That means the entire architecture has been reworked. From the Front end to the Compute Units, Caches etc.
It seems to me that Pascal is like Maxwell 3.0. Sort of like AMD going from GCN 1.1 to GCN 1.2. Arctic Islands is not GCN 1.3... it has undergone far too many changes that AMD has dubbed it Post GCN (it's a new ISA after-all). On a side note, I think that nVIDIAs post Maxwell architecture will be Volta.
When Nvidia announce Pascal after volta, it become pretty clear that Pascal is just a Maxwell with HBM. There is no reason for Nvidia to anounce a product(Volta) before the one they are going to release first (Pascal) unless it is a refresh. You guys can take a look at Haswell Refresh, Kabylake. Pretty much the same thing.
At the time Nvidia announce Pascal, Nvidia must be been thinking the result of HBM, node changes is enough to give Maxwell a significant performance gain, ontop of that with AMD remain uncompetitive at the time. That means Nvidia at that time think Pascal will be enough to complete with AMD.
A lot of people seems disagree with me that Maxwell is memory bottlenecked back then when 980 launched. IMO Maxwell could have been much faster if not for their narrow memory bus, take a look at GTX 960 vs GTX950, the latter have 25% shader cut off, slightly lower clockrates and GDDR5 being artificially limited to 6.6GHz, with all that 950 still is almost as fast as 960. 980Ti shares the same thing too, it has a couple of shader cut off, but it is still as fast as a fully enabled TITAN X. If GTX970 didnt have their memory bus cut off, it would have been as fast as 980 too.
Now imaging if all that memory bottleneck is being lift off by using HBM, as well as the power saving from HBM that allows the GPU to clock much higher. (may be 1.5GHz). So will Pascal be fast enough for 1 generation? More than enough, there is no need to bring new architecture Volta out on the table now on a immature 16nm node.