Intel unveils chip ideas at the cutting edge

Faster processors need a few tricks, and Intel still has plenty up its sleeve, as it revealed at its Developer Forum in San Francisco

Intel's push for faster, smarter, lower power chips shows no sign of stopping, according to its engineers.
In a press briefing before the start of the Spring 2004 Intel Developer Forum, the company showed off new ideas and demonstrations across a range of technology areas.

Among the goodies was a 64-bit arithmetic logic unit -- one of the basic building blocks of processors -- which was "the fastest so far, running 20 percent faster with 56 percent lower energy," according to Shenkar Borkar, director of Circuit Research at Intel -- "but we call him the professor," said Borkar. Figures showed it taking around 100 milliwatts at 0.9 volts and 2GHz, rising to two watts at 7GHz and 2.1 volts. Low power consumption for such ALUs is essential in modern processors, which can contain several "and the power saved soon adds up" said Borkar.

Another technique for reducing power and increasing performance has been christened swapped body biasing. Forward body biasing is an established way to reduce wasted leakage current through transistors -- one of the big power hogs in portable devices -- by electrically altering their characteristics, at the cost of reducing their performance. "So the idea came to us", said Borkar, "that by reversing the bias we could get better performance but at the cost of leakage." By swapping between the two types as different areas of the chip were used, the best of both worlds could be achieved, he reported. As an example, Borkar showed a network accelerator chip -- designed to work at 10GHz but running at much lower speeds and voltages -- that gave 60 percent better performance at 1GHz than the design without the swapped body biasing. It also reduced leakage current by two-thirds. This was an example of how a high speed, high power chip could be reused as a low power part, Borkar added.

High speed data transfer between chips is another perennial problem. "We have all these transistors on the chips," Borkar said, "but they're no good if you can't get data to them. Most buses between chips are simple-minded affairs, with the data represented by a succession of different voltages. By using some of the transistors on a chip to create high speed modems, performance even on ordinary motherboards can be vastly improved -- it's analogous to ADSL modems getting megabits across a phone line where ISDN only manages kilobits. Intel has achieved speeds of  8GBPs on standard motherboard construction with 130nm chips, which Borkar claimed as a record for a parallel data bus.

Another headache for chip designers is the limit to the number of memory chips that can be plugged into a motherboard -- too many at once and they overload the bus. Memory transfer speeds can be maintained for much larger numbers of chips by daisy-chaining them like fairy lights instead of plugging them all in parallel. This point-to-point system, developed by Intel in conjunction with Samsung and Infineon, relies on each chip to pass on signals to its neighbours: in another idea borrowed from telecommunications, two chips can be sending signals to each other on the same wire at the same time. The receiver in each chip listens to its own transmissions and cancels them out, leaving just the signals from the adjacent device.

Krishnamurthy Soumyanath, director of Communications Circuits Research, showed off some future radio circuit building blocks. Intel is very interested in bands at 60GHz and above, which offer very high bandwidths for localised wireless networking but are currently fallow due to the difficulties of making radios at those frequencies. There is a great deal of debate in the industry as to whether exotic materials such as gallium arsenide and indium phosphate are necessary -- the traditional approach -- or whether the same silicon designs that Intel champions can do the job.

As part of this debate, Soumyanath revealed tunable oscillators -- essential for both transmission and reception of radio -- that use the same class of transistors as currently found in today's 90nm processors, but run at speeds of up to 100GHz. The basic structure of these oscillators is very similar in concept and execution to traditional microwave circuits. The big difference is that where engineers have ordinarily laid out patterns of components using variants of ordinary wiring, Intel has done this directly on the chip. Just as shrinking digital circuits to 90nm makes them much faster, applying the same ideas to analogue microwave radio circuits makes them go at a much higher frequency. Although the resultant oscillators are extremely low power, the demonstration proves that you don't need exotic materials to reach the new bands at 60GHz, said Soumyanath.

Details of all of the above developments will be published at the IEEE International Solid-State Circuits Conference 2004, currently also underway in San Francisco.