Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
AMD promises 20 nm products--next year
AMD had previously promised a 20 nm successor to Beema and Mullins for next year, so that might be what they're talking about. It might also be the only product that they're talking about.
In particular, GPUs are still stuck at 28 nm for the remainder of this year if AMD's first 20 nm product doesn't ship until next year. Doesn't ship until next year means doesn't launch until far into next year, as it takes months between when AMD starts shipping a product and when there is adequate volume all over the world to do a product launch.
What about Nvidia, you might ask? AMD nearly always reaches a new process node several months before Nvidia. This is partially due to Nvidia emphasizing larger chips than AMD; it's easier to deliver a smaller chip on a given process node. If AMD isn't going to deliver 20 nm GPUs this year, it's unlikely that Nvidia will, either.
There were already other rumors that both AMD and Nvidia were preparing yet another generation of 28 nm GPUs. Now, there have been some process node improvements. TSMC now has an 28 nm HPM option, which is reportedly better than was available for the early 28 nm process nodes. Their 28 nm production is now very mature, which means that AMD and Nvidia have a much better idea of what chip designs will work in the real world than when designing the initial products. Global Foundries now has 28 nm process nodes out--which AMD already used for Beema and Mullins, so it's plausible that discrete GPUs could go there, though more likely for AMD than Nvidia.
Worse, it's still entirely possible that there will never be discrete 20 nm GPUs at all. The 20 nm process nodes at both TSMC and Global Foundries are claimed to be optimized for low-power SoCs. That's great for cell phones and tablets, and probably suitable for a Beema/Mullins successor, too. But it could easily be terrible for high-power discrete GPU chips.
That would probably leave us waiting for TSMC's 16 nm FinFET process node for the next round of process node-driven desktop GPU advances. TSMC hopes to have that shipping in large volumes next year, so it's plausible that we could see video cards on it launch next year. But of course, "hopes to" and "is plausible" are not at all the same thing as "is going to happen".
Comments
I'm surprisingly ok with this.
Right now there isn't much stressing the current generation of GPU tech out there. Software can't really use all the GPU power we've had in the last couple of generations; at least without stretching out past 1080/1200 to multiple monitors or 4K resolutions.
I think 4K will be the driving force for additional GPU power. As we start to see 4K prices drop (and they are coming down surprisingly fast) I think we'll see the need for that GPU power that a better process node will lend us (or rather most likely, the better power efficiency will let us pack more units on the same die size, resulting in a faster GPU).
As we start to see 4K become more mainstream, we'll see a greater need for additional GPU power. I think there is probably room at 28nm to squeeze another generation out (at the typical ~15% performance bump per generation), we've seen a lot of advances in GPU Boost/PowerTune that have been able to squeeze more performance out of cards and give us some power &/or thermally limited designs, rather than just frequency / silicon being the limiting factor.
A little bit of refinement in the APIs may give us that bump without the need for silicon - Mantle and DX12 both promise to do that, and it looks promising out the gate but we are still a long way from anything like main stream adoption yet.