Most chip companies would complain if you called them sharks, but in one way the analogy is apt: if they don't keep swimming, they'll die. Devices must go ever faster, or the market suffocates. Intel has its eyes set on processors in 2010 that run 100 times faster than the basic Pentium 4. But computers that work at that target speed of a million million instructions per second -- one tera IPS, or 1TIPS -- have to circumvent some basic laws of physics that are already making designers sweat.
Making chips go faster is a matter of making transistors smaller. Tinier transistors -- the electronic on/off switches from which all computer logic is built -- not only fit more processing onto a chip but switch faster because they have fewer electrons to move. However, they're also less capable of halting the current when they're off -- so power is lost through them all the time. They're more like dimmers than switches, said Borkar. The extreme smallness of the circuitry also means that quantum effects come into play, with electrons able to tunnel through barriers no matter how perfect the insulation.
The effects are dramatic. A Pentium III had around 10 million transistors with a 250nm architecture: each transistor leaked around a nanoamp -- one-billionth of an amp -- so the total leakage current was around 10 milliamps, or barely enough to light an LED. Compared to the current the chip took in normal operation, it was negligible. However, a 45nm transistor -- the sort likely to be in mainstream chips by 2010 -- leaks 1,000 times as much. If no changes were made to counter the effects of shrinkage, and Moore's Law were to be successful in increasing the number of transistors, then a processor chip with a billion 45nm transistors would leak 1,000 amps even when completely idle. At 16nm, which will be in prototype by 2010, a chip only a centimetre square would dissipate 1300 watts, more than an electric kettle, before it had made a single calculation.
The thermodynamics of cooling such a thing are hard enough, but the economics are impossible. If you design a computer costing £800 today, you have around £35 in your budget to put power in through a power supply and take it out through cooling. For that money, you can just about afford to move 100 watts through the system. So, even if you could make your incandescently hot chip, you could never afford to put it in a system that anyone would buy.
But if physics is the enemy, it also has the cure. Because you have so many transistors, they're effectively free -- you can use as many as you like without impairing the efficiency of the chip. Put two in line as a switch instead of just one, and you reduce the leakage current by five to 10 times. Use transistors to turn off power to entire blocks of logic when they're not being used, and you can get up to 1,000 times less wasted energy. Make the whole transistor electrically negative, and you get another two to 10 times reduction in leakage: it goes a lot slower, but you can always use more transistors to switch positive voltage in when you need the speed. Use thick insulation that has the electrical properties of much thinner layers -- High K dielectric, in the parlance -- and you can cut out the tunnelling losses.
There are other tricks you can play once you've got an infinite supply of free transistors. One of the rules of power loss is that if you halve the voltage on a circuit it only uses an eighth of the power. It also runs at half the speed, so can only cope with half the data -- but if you then add a second copy of the circuit it can process the other half. You've doubled the transistor count, but you're only losing a quarter of the power that you would have lost if you relied on just one circuit.
Playing around with voltages is a good way to cut your losses in other ways. Feed low-speed circuits with low voltages, and their leakage goes down. Vary the voltage on parts of the chip to match the amount of work they have to do, and you can get very efficient -- even more so if you vary the speed at which they work as well.
This idea scales up very well, and is the reason Intel and everyone else is committed to multi-core processors. Although the details are still secret, it's known that great efficiencies can be had if you constantly vary the voltage to each core to match their work rate, and you can also move intensive processes around the cores to keep each one's individual temperature down. Leakage goes up with high temperatures, avoid hot spots and you'll save power.
Other techniques long known to increase processing efficiency are also highly beneficial for power saving. If your chip is spending energy even if it's not doing any work, then keeping it busy means more bang for little extra thermal hassle. Large caches which prevent the chip from twiddling its thumbs while it waits for data to come in from outside make a great deal of sense in low-power designs, as does hyperthreading: if one thread halts for a memory transaction, there's another waiting to kick off immediately. A three thread processor effectively triples performance for the same power and thermal design as a single threader.
Finally, designing custom circuitry to do single, intensive tasks can be very effective. The example Borkar showed was a circuit that handled TCP network traffic -- it did the job 10 times faster for two watts than a general purpose processor could manage at 75. Things like voice recognition, video/audio coding and graphics processing are all areas ripe for a little special attention. If you make your special circuitry out of more exotic materials better suited for the purpose and combine it with the general purpose silicon chip on a multi-part die, you can fine-tune this even further -- this is the systems-on-a-chip (SOC) design methodology that looks the most promising path beyond the end of the decade.
With the combination of all of the above techniques, Borkar said, Intel is confident of reaching its 1TIPS target by 2010 without burning out.