Whither Broadwell?

Quizzical

Broadwell is, of course, Intel's next generation CPU, built on a 14 nm Tri-Gate process node.

If you believe Intel's Tick-Tock claims from several years ago, Broadwell should have been a 2013 product.  Of course, since Westmere, Sandy Bridge, Ivy Bridge, and Haswell all slipped into the next year, people stopped believing those claims a long time ago.  So Broadwell would, of course, be delayed somewhat into 2014, but then all would be well.  Right?

Well, 2014 is now more than halfway over, and Broadwell doesn't look like it's coming soon.  Rumors say that if there is going to be a 2014 launch at all, it's going to be with token quantities and low-performance dual core chips only.  Intel had previously been ambivalent about whether Broadwell would come to desktops at all, so they have long been signaling that this isn't going to be a huge deal for gamers.

For quad cores, the situation is worse.  Depending on which rumors you believe, Broadwell quad cores seem to be delayed far into 2015--and they're no lock to be on the market a year from now.  Or maybe there are no Broadwell quad cores, but 14 nm quad cores won't come until Sky Lake.  Or the Broadwell quad core is delayed so far that it's successor, Sky Lake will hit market first.  Ignoring Sky Lake delays, of course.

All of this isn't promising for Airmont Atom, either.  Remember when Intel was promising that they would speed up Atom shrinks to be faster than Moore's Law?  Does anyone really think that high-priced Broadwell delays are to make room for Atom chips that sell for 1/5 of the price?  Once it finally shows up at retail, the AMD A10 Micro-6700T could easily be the top dog in Windows tablets for a long time.

What's going on here?  Well, it seems that the transition from 22 nm to 14 nm has been bumpier than Intel hoped.  This surely isn't helped by such a large jump in process node size:  the last time Intel tried to decrease the process node size by more than 1/3 in a single jump, it was most naturally measured in micrometers, not nanometers.  The new product on the new process node was the 286.  Some of you weren't born yet.  It's been a while, and bigger jumps are harder.

But Intel isn't the only one having trouble.  AMD moved their CPU production to 32 nm with Llano way back in 2011.  Come 2014, they've only gone down to 28 nm, and rumors put their next CPU architecture still at 28 nm.  AMD GPUs got to 28 nm early in 2012, and are still stuck there.  For comparison, Intel's Ivy Bridge was on 22 nm in 2012.

There are 20 nm process nodes on the way, of course.  But there is some dispute about whether they'll be sensible for video cards, and there may or may not ever be desktop cards built on 20 nm.  Some rumors have both AMD and Nvidia releasing yet another generation of cards later this year at 28 nm--which surely wouldn't happen if 20 nm were ready and suitable for GPUs.

Is this the end of Moore's Law?  Not yet.  Crucial launched SSDs on a 16 nm process node earlier this year.  Now, NAND is easier to do than complex logic circuits like CPUs and GPUs need.  What might be going on here is that, as each new process node costs more than the last one, it begins making more sense to skip nodes.

TSMC jumped directly from 55 nm to 40 nm, the largest percentage jump it had done in many, many years.  Then they had just as big of a jump from 40 nm to 28 nm, canceling the promised 32 nm process node that was to come between them.  Everyone but Intel seems to be working on a comparably large jump from 28 nm to 20 nm, and even the 20 nm may be unimportant with only one option per foundry.  TSMC is more focused on 16 nm FinFET (promised available for volume production next year), and Global Foundries and Samsung are working toward 14 nm.

The upshot is that die shrinks are becoming less common than they used to be, but when they happen, it's a bigger jump.  That makes it a more important jump than a smaller shrink would be, but not necessarily more important than the historical norm from a decade or two ago.  When you're limited by power, you only get about half the gains from a given shrink as when you're limited by transistor count, and products have generally moved from the latter to the former over the years.