End Of The Line For Moore's Law?

Moore's Law in its original sense is dead. It died in 2004. The CTO of IBM gave the eulogy, noting that "Somewhere between 130 and 90 nanometers, we lost scalability." This means that - unlike the previous decades - shrinking the die no longer makes it faster or lower power. That is why we see no 4 GHz Pentiums. They hit the wall with the 3.8 GHz Pentium 4.

The reason has to do with the metal, not the silicon. At 130 nm, 75% of the speed and power is spent in the aluminum interconnects. It takes more time and power to send signals between the transistors than for the transistors to switch.

180 nm was the last, old-school CMOS process. Up until then, you shrank the 2D geometry. This made the transistors and the metal smaller without changing much in the vertical dimension other than making the gate thinner. Smaller devices and metal meant constant relative resistance of the metal between the transistors, lower capacitance of everything and lower on resistance of the FETs. Shrinking the part made it cheaper (less silicon), faster and lower power. Big win.

But at 130 nm, the metal wires looked like skyscrapers: relatively tall and thin. Now, you have to shrink the height of the metal as well as its width, or the skyscrapers start falling over. Oops. Now the relative on resistance *increases* as you shrink. Delay between transistors increases as well as power dissipated waiting for the signals to get there.

Result? Below 130 nm, chips get cheaper but *slower* and *hotter* relative to 130 nm and above. That is why we do not have 4+ GHz Pentiums to this day, after decades of increasing CPU clock speeds.

This is not for lack of desire or effort. Faster speeds sell, even if you do not need it. It is like horsepower in a car or Watts in a stereo amplifier. And many have tried. Going form aluminum to copper helps a little. Using halfnium oxide and other tricks to lower capacitance helps a little, too. But these are all one-trick-ponys. They do not scale, meaning that they do not help you with the next shrink.

Going to germaium or un-obtanium will not help. Again, the problem is in the metal interconnects, not the transistors. Copper is the best metal out there for conductivity, and only slightly better than aluminum, the workhorse of the silicon industry. To beat it you need something that amounts to thin film superconductivity at room temperature, still in the future.

Implications? The single thread CPU has hit the speed wall. So software can no longer be bailed out by the next generation of hardware. The only way to go faster is to go parallel. This is not a general solution. Nobody know how to make a C compiler that will take a single threaded problem and efficiently convert it to parallel processing with a speed up proportional to N processors. And this is after decades of trying.

Parallel processing works great if you have a parallel problem. Web farms and chip verification linux farms have been around for decades and work fine. But that is because their problems are inherently parallel, and the parallel approach is built from the ground up.

So, why is multi-core suddenly popular? Well, what else are you going to sell for your next generation of chips? Maybe with enough smoke and mirrors, you can claim that a 20 core chip at 1 GHz each is equivalent to a 20 GHz processor. Or not.

Dave Wyland