» Current | 2020 | 2019 | 2018 | Subscribe

Linley Newsletter

LeapMind Jumps on Binary Networks

May 12, 2020

Author: Linley Gwennap

Sometimes, smaller is better. LeapMind has developed a deep-learning accelerator (DLA) that delivers excellent performance per watt by focusing on 1-bit (binary) data. Operating on these tiny values greatly reduces the memory and power required to run a neural network, enabling the company’s Efficiera design to achieve 28 trillion operations per second (TOPS) per watt. The downside of such small values is their lack of precision, which can undermine the accuracy of the neural network’s output. But the startup has developed innovative approaches that retain much of the savings while boosting the output accuracy to nearly the same level as more-common 8-bit designs.

LeapMind plans to license the Efficiera intellectual property (IP) for customers to include in their own SoCs. The initial design scales from 1 to 12 TOPS of peak performance, depending on the configuration and process technology. The power draw ranges from about 80mW at the low end to 1W at the high end, making it well suited to low-power edge devices that perform vision-recognition tasks such as pose detection, face detection, and object recognition. LeapMind plans a beta release of the IP this month, with general availability in the fall.

Other companies support binary neural networks (BNNs) in hardware. Earlier this year, Apple reportedly paid about $200 million to acquire Xnor.ai, a startup that was developing a chip to implement BNNs. Lattice offers its SensAI software and RTL overlay to run BNNs on some of its FPGAs. More recently, Cadence, Ceva, and XMOS have added some binary support, but these designs also handle more-traditional INT8 networks, making them less area efficient. None of these competitors specifically addresses the accuracy shortcomings of BNNs.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article

Events

Linley Fall Processor Conference 2020
Coming October 28-29, 2020
Hyatt Regency, Santa Clara, CA
More Events »

Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »