Bringing Cost-Effective, On-Device AI to Home Appliances - Sparse Weights and Interactions Negate GPUs and Cloud Computing

GPU silicon and cloud computing infrastructure are too costly for mass markets devices like refrigerators and washers. By deploying compute resources only to the necessary parts of AI inference, Sparse AI allows product designers to practically incorporate new AI features like natural voice interfaces into their consumer offerings without breaking the bank or electric bill surprises.

- Sam Fok

Artificial Intelligence (AI) is changing every aspect of consumer product design. Manufacturers are discovering new ways to deploy AI into their products to differentiate in the market with new capabilities, improved efficiency, and reduced operating costs. However, the extreme processing requirements of AI have led manufacturers to implement AI either on-device using a high-cost, AI-capable chip, with a connection to cloud-based AI infrastructure, or both.

These approaches incur design and operating costs that have priced AI out of all but the highest-end products in markets such as white goods/home appliances and other consumer electronics applications. To make AI accessible to cost-sensitive mass markets, we need to bring costs down. This is where sparse AI technology can help. By optimizing AI inference processing by up to 100 times, sparse AI enables developers to implement complex, deep-learning-based AI models using low-cost AI MCU silicon without adversely impacting speed, efficiency, memory footprint, or performance. This article will explore sparse AI and how manufacturers can optimize AI inferencing to reduce cloudbased dependency and infrastructure or even implement powerful AI-based capabilities completely on-device.

SPARSE AI

AI models can be complex and require extensive processing resources and memory. In high-end AI systems, a specialized (and expensive) processor like a GPU runs AI model inferencing on-device. Alternatively, many systems take a cloud-based approach where data is collected on-device and sent to a server for processing.