Jetson, Coral, NPUs and the race to run models everywhere.

Why AI Is Moving to the Edge

Not every model belongs in a hyperscale data center. For robots, vehicles, cameras and wearables, shipping data to the cloud introduces latency, privacy concerns and dependency on solid connectivity. That’s why “edge AI” platforms are booming—putting serious inference horsepower into boards the size of a credit card. Jaycon+1

Running models locally trims round-trip delay to milliseconds, keeps raw sensor data on-device, and lets systems keep working even when the network drops.

NVIDIA Jetson: Tiny Boards, Big GPU Muscle

NVIDIA’s Jetson line is a staple for robotics, drones and embedded vision. The Jetson AGX Orin module tops many 2025 edge AI rankings, with up to roughly 275 TOPS of AI performance for high-end robotics and autonomous machines. Jaycon+1

At the maker and prototyping end, the Jetson Orin Nano Super Developer Kit delivers around 67 TOPS in a $249 board, with substantially higher memory bandwidth than earlier Nano kits—enough to run small chatbots, vision agents or navigation stacks on-device. The Verge+1

Google Coral: Edge TPU for Ultra-Low Power

Google’s Coral platform takes a different approach, using a small Edge TPU ASIC as a co-processor alongside an Arm-based SoC. The Coral Dev Board pairs an NXP i.MX 8M CPU with the Edge TPU accelerator, targeting low-power vision and sensor workloads. Google for Developers+1

Benchmarks show a single Edge TPU can deliver around 4 TOPS at about 2 TOPS per watt—making it well suited for always-on detection tasks in cameras, industrial sensors and smart home devices. Google for Developers+1

Smartphone and PC NPUs: Edge AI in Your Pocket

Modern phones and PCs now ship with built-in NPUs tuned for generative AI and on-device assistants. Apple’s current-generation Neural Engine, for example, is rated around 35 TOPS and is paired with higher memory bandwidth to support more demanding AI experiences on the iPhone 16 line. Competing NPUs from Intel, AMD and Qualcomm in “AI PCs” hover in the 38–50 TOPS range. CRN

For many consumer scenarios—photo enhancement, transcription, personal assistants—these NPUs will become the default AI runtime, with the cloud reserved for the heaviest tasks.

Choosing the Right Edge AI Platform

Selecting edge hardware is less about chasing the biggest TOPS number and more about balancing:

• Power envelope (battery vs mains), operating environment and thermal constraints.

• I/O and sensor needs (cameras, LiDAR, industrial buses, wireless).

• Software ecosystem: CUDA and ROS for Jetson, TensorFlow Lite and PyCoral for Edge TPU, or mobile frameworks for NPUs. Promwad+2Google for Developers

The most successful deployments treat edge devices as part of a larger system, with carefully planned model sizes, update pipelines and monitoring.

Closing Thoughts and Looking Forward

Edge AI is turning every device—from traffic cameras to tractors—into a potential inference node. Over the next few years, expect:

• More powerful yet efficient system-on-modules aimed at robotics and industrial automation.

• NPUs in nearly every mid-range and high-end phone, laptop and car infotainment system.

• New development stacks that make it easier to target multiple edge chips from a single codebase.

If data centers are the “brains” of the AI era, edge devices are its senses and reflexes—and the chips powering them are evolving just as fast.

Reference Sites

1. “Top 10 Edge AI Hardware for 2025” – Jaycon – https://www.jaycon.com/top-10-edge-ai-hardware-for-2025/

2. “Jetson Orin for Next-Gen Robotics” – NVIDIA – https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/

3. “Nvidia’s $249 dev kit promises cheap, small AI power” – The Verge – https://www.theverge.com/2024/12/17/24323450/nvidia-jetson-orin-nano-super-developer-kit-software-update-ai

4. “Dev Board Datasheet” – Google Coral – https://www.coral.ai/docs/dev-board/datasheet/

5. “Edge TPU Performance Benchmarks” – Google Coral – https://www.coral.ai/docs/edgetpu/benchmarks/

Author: Serge Boudreaux – AI Hardware Technologies, Montreal, Quebec
Co-Editor: Peter Jonathan Wilcheck – Miami, Florida

#edge AI, #Jetson Orin, #Google Coral, #Edge TPU, #NPUs, #on-device inference, #robotics hardware, #AI dev kits, #embedded vision, #dge computing