Deep Learning with Heterogenous Computing

Machine learning has existed since originating in the ‘60s but has exploded in the past decade as it integrates into everyday life. Guest Daryl Nees, director of Americas sales at Hailo Technologies LTD., explained to co-guest Peter Hsu, solutions engineer at Premio Inc., and Host Daniel Litwin how Hailo is at the forefront of artificial intelligence (AI) advances. As a semiconductor solutions provider, Hailo accelerates machine learning algorithms with neural networks. Through utilizing edge learning, they process data directly on a device or a nearby server. Data can be pre-processed at the edge of a network, reducing need for costly cloud service processing. Edge learning is used in many industries and is a crucial component to remaining relevant in today’s world. Nees explained, “Everybody I talk to across all these industries believe their AI plan and their AI productization roadmap is really a critical part of the future of their corporate organizations.” One real-world illustration is the surveillance industry. For example, when pixelated images of guns are identified by the AI, police are notified.However, none of this would be possible without power. Power performance is a key benefit for AI function. Hailo prioritizes this with application-specific integrated circuits (ASIC), such as Hailo 8, which is made for running AI-specific neural algorithms. Hailo 8 is significantly smaller in architecture when compared to compact graphics processing units while still maintaining its edge on memory. “Having the onboard memory… gives us a major advantage in latency and speed performance over versions that might have offboard memory,” Nees said. Hsu summarized: “The whole point of AI is to increase efficiency of what we do. It’s going to improve production, improve quality. The impact is only going to get greater.”