Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Where are the 20 nm GPUs?
Two years ago, Intel was producing 22 nm CPUs. Today, that's still their state of the art. Two years ago, TSMC was producing 28 nm GPUs. Today, that's still their state of the art. Two years ago, Global Foundries was producing 32 nm CPUs. Today, they've moved to 28 nm. I'm not sure if Samsung was producing 28 nm chips two years ago; at most, they've also moved from 32 nm to 28 nm in the meantime.
Normally, you get a full node die shrink about every two years. Of the four biggest foundries, we've seen one or two do a half-node shrink in the last two years and that's it. Why should you care about this? Because that's what drives Moore's Law--and hence long-term performance increases.
Today, AMD and Nvidia have 11 GPU chips between them in their modern lineups. But look at the desktop release dates on them (with the top bin of the GPU listed):
AMD Tahiti (Radeon HD 7970/Radeon R9 280X) January 2012
AMD Cape Verde (Radeon HD 7770/Radeon R7 250X) February 2012
AMD Pitcairn (Radeon HD 7870/Radeon R9 270X) March 2012
Nvidia GK104 (GeForce GTX 680/GTX 770) March 2012
Nvidia GK107 (GeForce GTX 650*) June 2012
Nvidia GK106 (GeForce GTX 660) September 2012
Nvidia GK110 (GeForce GTX Titan) February 2013
AMD Bonaire (Radeon HD 7790/Radeon R7 260X) March 2013
AMD Oland (Radeon R7 250) October 2013
AMD Hawaii (Radeon R9 290X) October 2013
Nvidia GM107 (GeForce GTX 650 Ti) February 2014
* The top desktop bin of GK107 was the GeForce GTX 650, but Nvidia didn't release that bin until long after the GeForce GT 640, which was a gimpy card basically released as a way to get rid of bad GPU chips that shouldn't go into laptops; I've listed the release date of the GT 640, as Nvidia didn't start selling the "good" bins of that die in desktops until later
Note that 5 of those 11 chips are more than two years old, and three others are well over a year old. Of the three remaining chips, one (Oland) probably released in laptops more than a year ago, but wasn't brought to desktops for a while, and the other two would never have existed if not for delays in the new generation of process nodes. GM107 in particular seems to be something of a test part in desktops and is very much priced not to sell.
In late 2011, I made a post remarking how unusual it was that AMD's Juniper chip (Radeon HD 5770/6770) had not only been on the market for two years, but had actually been a sensible purchase continuously for two years. I don't know if that was unprecedented in the entire history of GPUs, but it was certainly unprecedented since the transition to 3D in the late 1990s.
But now, Pitcairn and GK104 have done basically the same thing, and GK106 will probably join them when it hits its two-year anniversary. Cape Verde has fared better than that, even: not only has it been a sensible purchase for two years, but it has been the only sensible purchase in its price range pretty much that entire time due largely to Nvidia's inability or unwillingness to produce a competitive ~$100 budget gaming card.
The reason this didn't happen historically is that you'd get a new process node every year or so that was a half-node die shrink of the previous. The performance and efficiency gains from the new process node were so large that the previous generation cards were pretty quickly obsolete in the sense of no longer making sense to buy one new.
Process nodes still quickly make old cards obsolete, but since I built my current computer in October 2009, GPUs have moved from 40 nm to 28 nm--with no intermediate steps--and that's it. 28 nm quickly made 40 nm obsolete, and 20 nm will do the same to 28 nm, but we're getting a much larger time gap between new process nodes than is the historical norm.
Comments
20nm chips will soon be out. The next nodes are 16 and 14, and it looks like 16 will be mostly skipped in favor of 14.
edit: I think part of the reason 20nm took longer to come out is that it is a quite different process than 28. 20nm has local interconnect, which changes things quite a bit for implementation. You can't just simply shrink the shapes, it all has to be redone, including characterization.
16 and 14 are different yet again, going to FinFet type devices.
------------
2025: 48 years on the Net.
Every process node is quite a bit different from the previous, though the reasons why it differs vary from one transition to the next.
"The reason this didn't happen historically is that you'd get a new process node every year or so that was a half-node die shrink of the previous. "
But 20nm isn't just a half-node shrink. As I pointed out, it is different, not just a shrink. This probably explains why 20nm took longer to come out.
------------
2025: 48 years on the Net.
Well yes, every new node is different from the previous and not just doing exactly the same thing with smaller geometry. A different shrink years ago involved moving from aluminum interconnects to copper, or adopting silicon-on-insulator. The details vary with every transition, but the fact that there are changes each time does not. If it were simple to make the new node, they'd have skipped the previous one and just gone straight to the new one.
Now, 20 nm is a full node shrink from 28 nm, as opposed to a half node, so that means a typical time gap of 2 years instead of 1. But we're coming up on 2 1/2 years and still no 20 nm GPUs seem imminent.