Andes Technology, a leading Asia-based supplier of high-performance low-power compact 32/64-bit RISC-V CPU cores and a founding Platinum member of the RISC-V Foundation, and Montreal based AI startup Deeplite, Inc., the creators of Lightweight Intelligence™ making deep learning AI models smaller, faster and more energy efficient, today announced the results of their joint collaboration to deploy highly optimized deep learning models on Andes RISC-V CPU cores based on AndeStar™ V5 architecture .
The proliferation of smart devices like AI-enabled home assistants in recent years provides an ideal target platform for deploying highly compact deep learning models into daily life. These devices are designed to operate at both low power and low computation resources. To function effectively, a home assistant must be easy to use and respond to user requests in real-time. Today, due to the compute and power requirements of complex AI models, most smart devices must send user data and requests to the cloud to carry out AI processing then returning the results to the smart devices.
Andes and Deeplite teamed up to enable human-machine interfaces like home assistants, to operate locally with little to no cloud connectivity required. The scenario is an embedded solution where a home assistant “wakes up” when it detects a person via a small camera. The goal was to optimize a deep learning model running on Andes A25 and D25F that are the first commercial RISC-V cores with DSP SIMD ISA for low-cost edge AI applications. The team started with a MobileNet model trained on a Visual Wake Words (VWW) dataset that was 13MB in size. Using Deeplite’s hardware-aware optimization engine automatically found, trained and deployed a new model less than 188KB in size and with only a 1% drop in accuracy.
“We have more and more industry use cases where we see a need for embedded, optimized deep learning models running on our RISC-V cores such as A25 and D25F that have DSP instructions to accelerate deep learning algorithms,” said Dr. Charlie Su, CTO and Executive VP of Andes Technology. “Deeplite has provided a solution that can be leveraged both internally within Andes as well as for our customers to bring deep learning on Andes RISC-V CPU cores to resource-limited devices at the edge.”
“I am thrilled with the results of this collaboration! Not only has Deeplite delivered a 69x industry-changing deep learning optimization with minimal accuracy impact but we have done so by automating formerly manual techniques for neural architecture design that were time-consuming and error prone,” said Nick Romano, CEO of Deeplite, Inc. “What used to take weeks of expensive trial and error is now accomplished automatically in a few hours! Lightweight Intelligence™ by Deeplite and best of breed hardware from Andes are taking us one step closer to enabling AI in the things we use every day.”
By combining industry leading optimization by Deeplite with Andes’ state of the art hardware for use cases like voice recognition or person detection to meet microcontroller-level memory and compute requirements, device OEMs and application developers may offer users the benefit of keeping their data on-device, while still providing the real-time and seamless responses necessary for real-world AI everywhere.
Read More: Data Privacy as the New User Experience