''Something's gotta give'', as they say. Tough choices may lie ahead, as we balance the desire to continue AI development with our concerns about climate change. https://theconversation.com/it-takes-a-l...gry-151825
The good news there is that not only are current AI accelerators reducing their energy requirements per Moore's Law*, their different structure allows power savings options that you don't have on regular von Neumann machines (i.e. ordinary computers.) I expect it to be a problem for 10 years or so, with energy per GOP declining steadily during that time. (* - a law that says that that computer performance doubles every two years, computer cost halves every two years, with power consumption for both staying about the same.)
Moore's Law as applied to CMOS technologies does indeed have a limit. Once conductors and devices become less than about 1nm across, quantum effects start to dominate. We are now at 5nm. However, that's only for CMOS transistor technologies. Quantum computing is a very new field, but already we've seen orders of magnitude improvements in computational speed. But the computations that such computers make is very different - which means new programming languages and new ways of defining computing problems.
The tricky part of all of this comes down to how to get tech companies to maintain power consumption and price, while creating faster chips. Having said that, what would ''replace'' Moore's Law?
Several things will start to limit the ability to continue Moore's Law indefinitely, I think. The first is the speed of light. There will come a speed where a signal cannot get across the distance of the processor before it's needed. There's no way around that right now. As an example, right now very fast processors run at about 8GHz. At that clock speed, light can only travel about an inch before the next clock pulse. Since processor wafers are less than an inch across that works - for now. The second is Heisenberg (mentioned above.) Below a certain size, quantum uncertainty plays a larger and larger role in the output of a logic gate. This translates to spurious errors. Right now all computers see spurious errors but they are rare - 1 in 10^9 to 10^12. That is affected by temperature, radiation etc. When the errors become common it will make it difficult to design effective computers. The third is quantum decoherence. Quantum computers rely on quantum coherence to function; outside influences (from things like Brownian motion) can disrupt that. Hence the need to supercool many quantum computers.
