Microcomputers at ISSCC: Smaller Size, Lower Power Consumption, Stronger Computing Power

Source: CSDN

Computer scientist David Blaauw took a small plastic box from his bag and carefully picked up a small black cube with his fingernail, placing it on the hotel coffee table. This is one of the smallest computers in the world, measuring only 1 cubic millimeter. I had to be careful not to cough or sneeze, lest it be blown away or swept into the trash can.

Blaauw and his colleague Dennis Sylvester (both IEEE Fellows and computer science professors at the University of Michigan) presented ten papers related to these “microcomputers” at the IEEE International Solid-State Circuits Conference (ISSCC) held in San Francisco in February. They have been continuously developing various micro devices in recent years.

The overall goal of their Michigan Micro Mote (M3) program is to manufacture smarter and smaller sensors for medical devices and the Internet of Things, enabling them to do more with less energy consumption.

How to handle the data generated by a trillion devices

Microphones, cameras, and other sensors constitute the “eyes” and “ears” of smart devices. They must remain online in real-time and often send local data to the cloud, as they cannot perform complex analytical calculations on their own. It is predicted that by 2035, there will be a trillion such devices. Blaauw said, “If there are a trillion devices continuously generating data, humanity will be overwhelmed by data.” Blaauw and Sylvester hope to make these devices safer and more energy-efficient while occupying lower bandwidth by developing small, efficient computing sensors capable of performing analysis on board.

At ISSCC, they elaborated on microcomputers with power consumption of just a few nanowatts to perform tasks such as distinguishing the sounds of passing vehicles, measuring temperature, and light levels. They also demonstrated a compact radio that can transmit data from a small computer to a receiver 20 meters away, a significant improvement compared to last year’s reported range of 50 centimeters. They also introduced a collaboration with TSMC (Taiwan Semiconductor Manufacturing Company) to embed flash memory in devices and showcased dedicated low-power hardware for running AI algorithms based on deep neural networks.

Blaauw and Sylvester stated that they have taken a comprehensive set of measures to achieve these new features without increasing power consumption. Sylvester noted that there is no simple way to explain how they did it. If there is, it is “intelligent circuit design,” Blaauw added.

Sylvester said that memory research is a good example of how the right trade-offs can enhance performance. Previous microcomputers used 8 kilobytes of SRAM (Static RAM), which inevitably resulted in poor computer performance. To record video and sound, microcomputers need more memory. So the team collaborated with TSMC to embed flash memory in the circuitry. Now they can manufacture small computers with 1 megabyte of memory.

Flash can occupy less space than SRAM while storing more data, but writing to memory requires a lot of energy. Their team, along with TSMC, designed a new memory array that uses a more efficient charge pump for the writing process. This memory array has slightly lower density than TSMC’s commercial products but is still much stronger than SRAM. Sylvester said, “We achieved significant gains through small trade-offs.”

Neural network processor with a power consumption of only 288 microwatts

Another microcomputer they proposed at ISSCC contains a deep learning processor that can perform neural network calculations using only 288 microwatts of power. Neural networks are AI algorithms that perform well in tasks such as facial and voice recognition. They typically require a lot of memory and powerful processing capabilities, so they usually run on server clusters equipped with high-end GPUs. Some researchers have been trying to reduce the memory and power consumption requirements of deep learning by using dedicated hardware designed to run these algorithms. Even so, these processors still consume more than 50 milliwatts of power—which is still too much for microcomputers. The University of Michigan team reduced power requirements by redesigning chip architecture, for example, by placing four processing units in memory (in this case, SRAM) to minimize data movement.

The current idea is to embed neural networks into the Internet of Things. Blaauw said, “Many motion detection cameras capture the branches of trees swaying in the wind, which is meaningless.” Security cameras and other connected devices are not intelligent enough to distinguish between thieves and trees, so they waste energy sending meaningless images to the cloud for analysis. On-board deep learning processors can make better decisions, provided they do not use too much power. The University of Michigan team believes that deep learning processors can be integrated into many other IoT applications besides security systems. For example, if a central air conditioning system sees multiple people putting on coats, it can decide to turn off the air conditioning.

After demonstrating various applications of these microcomputers in academia, the University of Michigan team hopes to bring products to market within a few years. Blaauw and Sylvester said their startup CubeWorks is currently prototyping and conducting market research. The company quietly launched at the end of 2013. Last October, Intel Capital announced an undisclosed investment in the company.

Click to read the original text for more