Accelerating Edge AI Development in Real-Time Control Apps

Members can download this article in PDF format.

Much of the electronics world today obsesses over artificial intelligence (AI) in big system-on-chips (SoCs) and complex, multichip designs. But what may be overlooked are far-reaching applications that can be enabled inside smaller chips like microcontrollers, making both industrial devices and consumer devices more intelligent and more efficient.

The reason? Primarily response times, power consumption, performance, development complexity, memory footprint, and the cost to transform data into the real-time decisions required with AI capabilities. Fortunately, there are strong indicators that’s all changing.

Sponsored Resources:

Consider Texas Instruments’ MCUs with integrated TinyEngine neural processing units (NPUs). These dedicated hardware accelerators optimize deep-learning inference operations to reduce latency and improve energy efficiency when processing at the AI edge. With this innovation, engineers are now able to deploy intelligence almost anywhere.

The idea behind the TinyEngine NPU is to execute computations required by neural networks in parallel to the primary CPU running application code. It addresses key design constraints that have prevented widespread adoption of embedded AI by utilizing 120X less energy per inference and 90X lower latency compared to software-based AI.

By executing machine-learning algorithms in parallel to the primary CPU, real-time processing of neural-network models can take place on resource-constrained devices. This optimizes deep-learning inference latency and power consumption when processing at the edge, eliminating the round-trip latency of cloud-based inferencing for increased system responsiveness.

TI plans to integrate the TinyEngine NPU across its entire microcontroller portfolio (Fig. 1). It will lead to expansion of edge AI capabilities into devices such as portable battery-powered products, medical wearables, as well as personal electronics and industrial equipment that were previously unable to support meaningful AI workloads.