There is a lot of focus on the end of Moore's law here, but it isn't main driver of what's happening. The slow down we are seeing is driven by the end of Denna

There is a lot of focus on the end of Moore's law here, but it isn't main driver of what's happening. The slow down we are seeing is driven by the end of Dennard scaling. There is a thermal dissipation limit of 200W per chip for air cooling. We hit that decades ago of course, but it didn't matter while Dennard scaling kept dropping the power consumption. Once that stopped we squeezed a bit more of stuff by being more power efficient, which boiled down to two things - turning stuff off when it wasn't needed and devoting the transistors Moore's law gave us to specialised tasks (like silicon dedicated to encryption or h264 encoding) that did the job more efficiently. However, that doesn't get you very far. Which is probably why he didn't mention 3D, even though we have it 32 layers of it now and they are talking about 256 layers. What is the point of having 128 CPU's on a single die just running 4 of them exceeds your power budget? Indeed, what's the point of spending billions perusing Moore's further? Or to put it another way again, the human brain fits roughly the same number of synapses per unit volume as modern 3D silicon has transistors. The brain's raw switching speed is roughly 1,000,000 times slower than silicon (1ms vs 1ns), but power consumption of a synapse vs a transistor is roughly 100,000 times better. So while AlphaZero learnt to play Go between than any human in a few days, it used more energy that an entire human (not just their brain) would use in several life times to do it.