"Stanford Engineers Present New Chip That Ramps up AI Computing Efficiency"
Edge computing powered by Artificial Intelligence (AI) is already pervasive as drones, smart wearables, and industrial Internet of Things (IoT) sensors contain AI-enabled chips. This enables computing to take place at the "edge" of the Internet, where the data originates, allowing for real-time processing while also ensuring data privacy. However, the energy provided by a battery limits the AI capabilities of these tiny edge devices, so increasing energy efficiency is critical. Data processing and storage occur in separate units in today's AI chips: a compute unit and a memory unit. Since frequent data movement between these units consumes most of the energy during AI processing, reducing data movement is critical to addressing the energy issue. Engineers at Stanford University have proposed a novel Resistive Random-Access Memory (RRAM) chip as a possible solution that performs AI processing within the memory itself, eliminating the need for a separate compute and memory unit. Their "compute-in-memory" (CIM) chip, NeuRRAM, is about the size of a fingertip and can do more with less battery power than current chips. Performing those calculations on the chip rather than sending data to and from the cloud could lead to faster, more secure, cheaper, and scalable AI in the future, giving more people access to AI power. While CIM has been around for decades, this chip is the first to demonstrate a wide range of AI applications on hardware rather than just through simulation. NeuRRAM is currently a physical proof-of-concept (POC), but it requires further development before it can be translated into actual edge devices. The chip's combined efficiency, accuracy, and ability to perform multiple tasks demonstrates its potential. This article continues to discuss the architecture, possible impact, and potential applications of the NeuRRAM chip.