For 50 years, semiconductor makers have been doubling the number of transistors per unit area on chips every two years or so, following Moore’s Law. But if that continues for a few more years, transistors will be no bigger than a few atoms, a size where classical electronics becomes unreliable and quantum weirdness reigns. Semiconductor development needs to take off in new directions, and IEEE is helping by producing the road map.
IEEE began work earlier this year on the International Roadmap of Devices and Systems (IRDS) initiative, an IEEE Standards Association (IEEE-SA) program sponsored by the IEEE Rebooting Computing initiative in consultation with the IEEE Computer Society. The document is designed to replace the International Technology Roadmap for Semiconductors, which the semiconductor industry developed and periodically revised since the 1990s.
The IRDS isn’t much like most maps of roads and highways, because the destination points aren’t completely known and the highways haven’t been laid out yet. The IRDS will provide general directions only, giving designers and manufacturers insight into the most likely technical events during the next 15 years and a way of tracking their progress. It also can help them identify upcoming kinks, bottlenecks, and roadblocks that need to be fixed to avoid real headaches, and it can compare potential solutions.
“It’s a big picture,” says IEEE Fellow Paolo Gargini, who chairs the IRDS initiative, “and researchers have to fill in the fine details.”
As the word systems in its title indicates, the IRDS will give an even bigger picture than its predecessor. Instead of building smaller, faster, cheaper devices that people then find uses for, the new initiative will start with existing applications and the kinds of devices best suited for them, connecting the people who want to build different kinds of equipment with the people who’ll make the chips that will go into them.
The IRDS’s application-oriented approach could reshape the semiconductor industry’s economics. Instead of a few chip types sold by the millions, there could be many chip types, each focused on particular application sets.
So far, says IEEE Fellow Tom Conte, who chairs one of the IRDS focus teams, the group has been concentrating on transistors for microprocessors and for memory. But not everything requires those kinds of devices.
“For example,” Conte says, “neuromorphic computing, which applies to our admittedly incomplete knowledge of the brain, requires very different devices from the ones we now have—analog processors, resistive memory that can store multiple values—and we want to integrate that onto one chip.”
The IRDS focus groups will cover a variety of application areas including big data analytics, facial and image recognition, event and physical-system simulation, optimization (quantum computing has a huge potential there, Conte predicts), graphics rendering, media processing, cryptography, and wireless communication. The goal, Conte says, is to take the applications and “drive them down to the best devices to build and the best systems we need.”
Instead of requirements coming down to “just build smaller and faster,” people are looking at cutting-edge ways to compute or do memory.
Not all applications require high precision. Human perceptions can tolerate a lot lower precision than computers now deliver. When computers interface with humans, figuring out when “just good enough” actually is good enough and computing only to that level can lead to greater efficiencies.
Not all the paths in the new road map will be for specific applications. Trying to match the progress rate of the past 50 years will require new physical designs such as multiple-layer 3-D chips, new materials such as graphene, and new architectures such as uniting memory and processing on the same chip, to enhance speed while reducing heat production.
Some focus groups are likely to deal with physical issues such as chip packaging and production. Others will cover metrology, testing, and how to improve production yields.
And not all the ideas to be explored will be new. Some will be approaches that were shelved when following Moore’s Law was still the easiest way to go.