Elevating Edge AI: Unlocking Unprecedented Performance and Power Efficiency

The evolution of technology is a familiar story, and artificial intelligence is no exception. Each new generation of AI chips adds more MACs, deeper quantization, and additional features—all in pursuit of higher TOPS/Watt. Yet, as users’ expertise grows, the demand shifts from simple usability to unmatched algorithmic flexibility, maximum throughput, and minimal power consumption.

Through the CEVA NeuPro platform, we have successfully introduced edge‑AI solutions across diverse markets. Today’s users want more than incremental upgrades; they require revolutionary improvements that enable them to outpace the state‑of‑the‑art by an order of magnitude.

Measuring Real‑World Impact

While TOPS/Watt is a convenient benchmark, it falls short in real applications. For visual inference, frames per second per watt (FPS/W) is far more meaningful. In automotive safety, for instance, a perception engine must process at least 100 FPS with the lowest possible power draw, as delays can mean the difference between a safe stop and an accident. Achieving such performance demands far higher FPS/W than conventional metrics suggest.

The market opportunity is clear. Automotive and telecommunications are poised to drive the largest share of edge‑AI growth, with intelligent imaging already dominating automotive and the “many‑camera” trend reshaping mobile phones. Modern smartphone cameras are replacing legacy algorithms with neural networks for denoising, stabilization, and super‑resolution, all running at 60 fps within tight energy envelopes.

What a Major Advance Requires

While analog AI and spiking neural networks offer intriguing possibilities, product makers still need scalable solutions that can hit volume today. This leaves a wealth of algorithmic potential: extensive quantization options, Winograd support, sparsity optimizations, diverse data‑type handling across varied bit‑sizes, parallel vector processing, data compression, matrix decomposition, and support for next‑generation architectures such as transformers and 3D convolutions. These capabilities can deliver up to a 50:1 acceleration over reference networks.

A Call to Action

Experienced AI developers now know precisely what they need to build. What they require is a platform that provides all the neural‑network components and optimizations they trust, enabling them to craft the optimal solution for their product.

This is more than a wish list; it is a roadmap for the breakthrough performance, throughput, and low power that advanced edge AI demands. Manufacturers can no longer settle for incremental gains—they expect platforms that match their deep understanding of AI possibilities.

To learn more about CEVA’s work in edge AI, click here.

Roni Sadeh has over 20 years of experience in processor and accelerator design, focusing on AI‑related software and hardware solutions for audio, speech, and computer vision. He is developing the next generation of AI accelerators scalable to hundreds of TOPs.

Roni holds a B.Sc. in Aeronautical Engineering from Technion University.