Simple Is Better Than Big Clunking Chips says Intel
Interesting to see Intel, for so long the exponent of the big clunking chip, argue that simple cores are better than complex ones.
Intel Fellow Shekhar Borkar told last month’s DAC that, instead of using a billion transistor chip to integrate ten cores each containing one million transistors, it would be better to integrate 100 cores of 100,000 transistors each or 1,000 cores containing 1,000 transistors each. It’s interesting because the academics have been saying this for years, while the industry kept cranking frequencies on more and more complex chips, as the only route to better performance. “Von Neumann is a poor use of scaling, all the energy is going on the communication between the processor and the memory. It’s much better to use 20 microprocessors running at 100MHz than one at 2GHz,” IMEC co-founder, Professor Hugo De Man, told the 2005 MEDEA+ conference. PicoChip, the wireless fabless semiconductor company, reckons the most efficient core is a 16-bit processor with three-way LIW, three-deep pipeline, requiring 1m transistors and fitting onto half a square mm of silicon at 90nm. Stepped and repeated across the die, this makes up a PicoChip multi-processor chip. Used in embedded applications the performance of a multi-core improves linearly with the number of cores. Used in a general purpose application multi-core soon runs out of steam. “It’s easy to get linear improvement in performance in one application”, says Rupert Baines, vp of marketing at PicoChip, “but it’s not easy in a PC where code is whatever the user wants to use at the time and there’s all the legacy software to support.” Adding cores to a general purpose processor results in fast diminishing benefits. “A second core improves performance 50 per cent, a third core by 30 per cent and, with six cores, you’re flatlining”, says Baines. While Alan Gatherer, CTO for communication infrastructure at Texas Instruments, reckons: “I’m not sure anyone knows how to build a generic multi-core architecture. It’s a great goal, but the chances of failure are 100 per cent.”Tags: fabless semiconductor company, flatlining, Imec, Intel, legacy software, linear improvement, picochip, texas instruments