High(er) Frequency means that you have to run at a higher core voltage (this is, in part, because gate propagation delay is inversely proportional to bias voltage, i.e., the voltage corresponding to a 'high' or 1 bit). You have to decrease the propagation delay so that the clock tick gets everywhere in the processor quickly (one part of the clock doesn't lag the other parts, called clock skew). So in order for things to run faster, you'd have to run it at a higher V(bias). Now that means that there are higher thermal costs (things get hotter) - heat produced is proportion to voltage.
So its now mostly a thermal management problem. This is the primary problem in the newer chips. Even though we can pack in more transistors, we can't get signals among them faster without higher V and making it run too hot.
Therefore, our solution is to use the extra transistors to create a separate new processor, running with a different clock, so the 'tick' doesn't have to reach all parts of this rather big chip, but just needs to be synchronized intra-core.
I'm curious if asynchronous logic will get brought back in as time goes on. Multi-core is essentially desynchronising the individual cores since (iirc) clock generation and distribution is one of the most expensive parts of the processor, in both energy and size.
Historically, it lost out because of the added complexity of handshaking/synchroniser logic scattered all over the place, and the mess a mishandled metastable condition propagating through the system could cause. How that we've got transistors to burn, however, and routing is no longer done on large empty floors with crepe tape...
Well, Handshake did make an asynchronous ARM core not too long ago, and Achronix has a nice business in async FPGAs, so the idea isn't dead. That said, I'm not sure (though I'd love to be proved wrong) that the idea will really go big again. Synchronous chips are still easier to reason about.
By far the most expensive real estate on a typical processor is on-chip RAM, and that really does need a clock. Sure, clock-tree synthesis is complex enough that you may even still be able to start up a company selling a tool for it, but it's still possible for a few reasonably competent engineers to do within some weeks.
Right, but decreasing the size of the chip means that you effectively can get signals around it faster (because they don't have as far to go) and allows them to be run at lower voltages as well - modern CPUs seem to be generally around 1V, but 10 or 20 years ago they were around 2V, so there is (or has been at least?) a reduction in that over time as well.
So its now mostly a thermal management problem. This is the primary problem in the newer chips. Even though we can pack in more transistors, we can't get signals among them faster without higher V and making it run too hot.
Therefore, our solution is to use the extra transistors to create a separate new processor, running with a different clock, so the 'tick' doesn't have to reach all parts of this rather big chip, but just needs to be synchronized intra-core.