Hybrid Chip Containing Processors And Memory Runs AI On Smart Devices

A group of researchers from Stanford have developed a way to combine processors and memory on multiple hybrid chips to allow AI to run on battery-powered devices such as smartphones and tablets. The team believes that all manner of battery-power electronics would be smarter if they could run AI algorithms. The problem is efforts to build AI-capable chips for mobile devices have run up against something known as the "memory wall."

The memory wall is the name for the separation of data processing and memory chips that have to work together to meet the computational demands of AI. Computer scientist Subhasish Mitra says the transactions between processors and memory can consume 95 percent of the energy needed to perform machine learning and AI, severely limiting battery life. Stanford researchers have designed a system that's able to run AI tasks faster with less energy requirements.

They were able to do this by harnessing eight hybrid chips that each have their own data processor built right next to its own memory storage. The new study builds on prior work from the team on a new memory technology called RRAM able to store data even when power is switched off with more speed and energy efficiency than flash memory. RRAM allowed the team to develop a hybrid chip previously, but the latest design has a new critical element.

The new critical element is algorithms allowing the eight separate hybrid chips to be melded into a single energy-efficient AI-processing engine. The team says they were able to "trick" the eight individual hybrid chips into thinking they're one single chip, leading the system to be called the "Illusion System." The team says that the eight-chip system is only the beginning.

Simulations have shown that systems with 64 hybrid chips can run AI applications seven times faster than current processors while requiring only one-seventh as much energy. Researchers believe the prototype system's performance shows they're on the right track and that Illusion Systems could be market ready within 3 to 5 years.