The breathtaking advances in digital integrated circuits over the past 40 years—with current chips topping one billion transistors—have been possible in large part because designers and users have been able to ignore a number of “second-order” aspects of circuit behavior. Unfortunately, as process geometries continue to shrink, some of those second-order effects have been promoted to first-order headaches.
Take leakage current. It used to be one of those second-order effects that most digital chip designers and users could ignore, but now it’s often the dominant type of power consumption. Worse yet, it varies significantly with variations in fab process and operating temperature. So, say your leakage current eats 50% of a chip’s overall power, and this current varies by up to 300% depending on how the fab was running at the time it was manufactured, and on the chip’s temperature. As a result, total power consumption may vary by a large amount from one sample of the chip to the next—and is extremely difficult to model or predict. This may be a huge problem for system designers, especially since optimizing for power consumption has become such a critical part of many designs.
Then there’s electromigration, which is becoming more problematic as process geometries shrink and current densities in metal interconnects rise. J.R. Lloyd has a great tutorial on the physics involved, but the practical significance of electromigration is that it can degrade metal interconnects, increase resistivity (which reduces performance) and ultimately lead to opens or shorts in metal interconnect that result in catastrophic chip failure. When wires are relatively thick and current densities are low, this effect can be safely ignored—or at least, effectively counteracted. But down in the 65 nm region, electromigration is becoming harder to contend with.
For processor vendors and users, electromigration has an interesting implication: a processor’s lifespan may depend on how aggressively you use it—just like a car. You can imagine a for-sale ad that says, “Used iPhone, used only on Sundays by little old lady for calling grandkids, never ran H.264.”
Moore’s Law isn’t dying, but we’re going to pay stiffer penalties to keep it going. System designers and chip users will be forced to make new trade-offs in terms of performance versus longevity and reliability. All those niggly, second-order effects whose description we slept through in EE 101 are going to come back to haunt us. And as usual, there are unlikely to be easy solutions.
Jennifer Eyre White of BDTI contributed to this column.
Add new comment