There’s always something interesting on EE Times, and just the other day that something was a roundup by Rick Merritt of what the speakers at last week’s Imec Technology Forum had to say. (Imec is an R&D center focused on nanoelectronics and digital technologies.)
My interest began with the headline – 12 Views on the Future of Electronics – and was really piqued by the first of those 12 views: “Neural processors overtake CPUs, GPUs.” Well, not so fast on CPUs being overtaken, but it’s always interesting to see what the industry thinkers and doers are up to.
Neural processors overtake CPUs, GPUs
Kinam Kim heads up Samsung’s semiconductor business. His comments (IMHO) are colored by his business. Nonetheless, it’s worth hearing what he has to say:
Kim predicted that the DRAM and NAND technologies that Samsung dominates will continue to evolve. But he suggested that neural-network processors will disrupt the CPU and GPU architectures of rivals Intel and Nvidia.
Microsoft builds AI plug-ins
Peter Lee, who heads up Microsoft Research, quite naturally used Microsoft’s Powerpoint to illustrate his talk. But he was using “an early version of Powerpoint that supports real-time transcription. Halfway through his talk, the feed shifted to automatically translate his comments into Dutch.” He predicts that AI will be built into all sorts of applications.
Sprinkling machine learning everywhere
Imec’s Praveen Raghavan spoke about an emerging capability that will let neural networks learn “in the field.”
To get there, Raghavan and colleagues have defined what they believe is the most cost- and power-efficient architecture for machine learning (below). It uses separate arrays of XNOR transistors and MRAM cells, storing one bit in each cell of the array.
While not using neural nets as such, the chip performs similar kinds of tasks. The trade-off is that the one-bit math it uses is less precise than most neural nets, requiring more arrays with more hierarchies, he said.
The stimulating mysteries of the brain
Tim Denison is the VP of research at Medtronic. He reported on his work on an implanted 120-Hz stimulator used to help those with Parkinson’s control their tremors. Although the implants are proving to be effective, Denison points out that researchers still don’t understand how the brain works. So they’re using the implants as a tool to find out more. “The work is part of an evolving field of neural prosthetics that ‘we think of as reconstructing the nervous system.’” All very promising, as is the fact that researchers exploring multiple different diseases are using cloud services and shared toolkits so that findings are available in near real-time to those working in similar areas.
Photonic chips shrink medical test gear
Another Imec researcher, Paru Deshpande, is working on creating photonincs devices in silicon. “The chips aim to replace relatively bulky, expensive optical components in today’s desktop-sized medical test systems so they can evolve into tomorrow’s handheld units.”
Packing smarts in wearable patches
Wearables are much in the news these days. Mario Konijnenburg showed off the 55-nm third gen MuselC, which “is the first in the family to incorporate flash (384 Kbytes), support for Bluetooth Low Energy, a Cortex M4 class processor, and encryption.”
We’ll be back with the next six futures in two weeks time.