The International Technology Roadmap for Semiconductors plots an expanding gap between the growth of processor performance and the lagging ability of I/O to keep up. Similarly, data sent over the Internet is rising by a factor of 100 every 10-12 years — about the pace of growth of processor performance — but the links between processors and the optical backbone of the Net are not expanding at the same rate.
“A byte-per-flop gap is opening up that’s a major limit on architectures that gets worse and worse as we head to 2020 and beyond,” said David Miller, an optical research lab at Stanford, speaking in an evening panel at the International Solid-State Circuits Conference.
“We are getting close to the point where optical is more attractive,” said John Stonick, a Synopsys scientist and chair of the panel. “But optical interconnects have always been the next thing, and the big questions are when and how we get there,” he said.
Keishi Ohashi, an optical expert from NEC Corp., said optical may follow a path like hard disk technologies. It took nearly twenty years for powerful magneto-resistive heads to emerge from their beginnings in labor-intensive magnetic coils, he said.
Bert Offrein, an optical expert with IBM Research, noted milestones and challenges IBM has seen. In 2008, the company built Roadrunner, the first petaflop computer, using optical interconnects to each server board.
Last year IBM created its Power P775 high-end server that brought optical links to the processor. It fused 56 fibre cables to modules with 56 transceivers integrated with help from Avago on to each IBM CPU.
“It required 100 additional assembly steps to bring the optics to the chips in addition to building transceivers themselves,” said Offrein. “That’s justified for some high-performance systems, but for general servers we need something easier,” he said.
Board makers are experimenting with many ways to integrate optical links, prototyping 10 Gbit/second channels now, but none