Moore’s Law is over. It was never a physical law. It’s an observation due at least partly to economics.

Problems with Moore’s Law appeared a few years ago with Nvidia’s 40nm to 28nm transition. Nvidia put out a presentation outlining issues they had with their foundry, TSMC. Particularly the cost per transistor did not decrease with the 40nm to 28nm transistion. This was partly because 28nm uses double patterning lithography. Double patterning is needed because extreme ultraviolet (EUV) lithography has faced setbacks & delays.

This was the first time that shrinking the process node yielded no economic benefit. The implications are obvious. There is less economic incentive for switching to or developing new process nodes.

Intel’s recent announcement of a ~1 year delay for their 10nm architecture to 2017 also indicates that Moore’s Law is at least severely strained. Intel’s current 14nm architecture, Broadwell, was also delayed by 6 months.

  • 65nm (Core), 3/2006
  • 45nm (Penryn), 11/2007
  • 32nm (Westmere), 1/2010
  • 22nm (Ivybridge), 4/2012
  • 14nm (Broadwell), 10/2014
  • 10nm (Cannonlake), 6/2017

It’s difficult to tell from the graph but it is no longer linear.