If I was an Intel engineer I would buy a Ferrari.
But I like my Honda Civic.
Basically what it comes down to is heat. Heat is the problem. Our former CTO (and now the VP of my group) Pat Gelsinger gave a great keynote talk on power at ISSCC back in 2001 or so. You can tackle heat at the architectural/microarchitectural level - I know ARM have plenty of papers about this out there. But it's a problem that crosses all levels of the design from architecture to logic to circuitry to process technology.
Each process generation is getting successively harder to design with. People have been stating that the end of "CMOS" is coming. They've been saying it even since 2um (2000nm) technology. But while, I don't see the end in sight, I do see that things are getting much, much harder than they used to be. More architectural complexity to achieve smaller performance gains; increasing problems with noise, heat, power delivery; increasing design team sizes; and lots of issues scaling beyond where we are using traditional silicon processing techniques.
But of all of these problem, power/heat are probably the two biggest. Transistors can be thought of as the equivalent of electronic water valves. You use one signal (the gate) to control water (charge carrier) flow from one pipe (the drain to the source, or visa-versa depending on whether you are a physist or an engineer). If you want to make the analogy a bit closer, imagine that you are using water pressure to adjust the valve.
Back around 0.5um or so, there was essentially no leakage worth speaking of in a transistor. Your water valve had a leak-proof seal (subthreshold leakage), your pipes didn't leak (substrate leakage) and the water pressure that turned the valve on and off didn't leak either (gate leakage). Now in 90nm, all three of these are leaking quite a bit more. The valve never fully shuts off, so water is always flowing through the valve. The pipes leak a bit, so water is falling on the floor. And the water that controls the valve leaks through the valve. That's a lot of leaking water (power).
Can this leakage be fixed? Can you plug the leaks? The answer is, of course, yes - mostly. But if you do, you sacrifice speed of turning on the valve. Similar to the real world, if you tighten things down, using thicker rubber and better seals, then switching the valve on and off is going to slow down. Leaving behind this analogy and back to the real world, about 35% of total chip power on 90nm designs is from leakage (shooting from the hip guess based on numbers I heard thrown about as ISSCC). So if you had a 100W part, if you could get rid of leakage it would be 65W. But the performance sacrifice is too much to consider ever doing this.
And the problem doesn't just affect Intel. Look at graphics chips that require two molex connectors in addition to the motherboard power. Look at the power of high performance CPU designs from other companies (IBM, AMD, even Sun, although they are lower than the rest). Power supplies with 12V dual-rails. Enormous heatsinks, loud fans. Extremely complex power delivery systems getting harder and harder as voltages continue to get lower, while current is rising. (100W at 1.5V is 66A, 100W at 1V is 100A). Meanwhile mainstream consumer preferences are towards smaller, quieter, less intrusive.
There are plenty of tricks that can be used to lower power. They add complexity to the design and can introduce their own problems, but that is where the future is headed.
Facebook
Twitter