Tech Manufacturers Unveil Energy-Efficient Chipsets for Edge Devices

Global shipments of energy-efficient edge AI chips are forecast to jump from 380 million units in 2024 to 1.2 billion by 2027, a 45 % compound annual growth rate, according to new data from IDC. The surge is being driven by factory-automation upgrades, retail-analytics roll-outs and U.S. infrastructure grants that require on-device processing for privacy and resilience.

One architecture gaining traction couples ferroelectric transistors (FeFET) with in-memory compute, eliminating the energy-hungry shuttling of data between separate memory and logic dies. A design unveiled in May by Bosch, Fraunhofer IPMS and GlobalFoundries delivers 885 TOPS/W—roughly 44× the efficiency of today’s GPU-based AI cards—and is now in pilot production for automotive camera modules. “The chip executes MobileNet-v2 at 1,200 frames per second while drawing only 8 mW,” said Thomas Kämpfe, Fraunhofer’s neuromorphic-systems group manager. “That lets a standard 12-V car battery power always-on vision for more than a month.”

Commercial off-the-shelf options are reaching similar milestones. Hailo’s latest Hailo-8™ M.2 accelerator hits 26 TOPS at 2.5 W (10 TOPS/W), while Axelera’s Metis AIPU scales to 214 TOPS at 15 TOPS/W, enough to run real-time transformer networks on 16 video streams inside a single edge server. AMD, fresh from its acquisition of low-power start-up Untether AI, previewed the Ryzen™ Edge X800 SoC last week: the 6-nm device combines eight Zen-4 cores with a 28-TOPS NPU that consumes just 1 W, enabling retail checkout kiosks to process vision and natural-language queries without a cooling fan.

“Power efficiency is the new battleground,” said Lydia Chen, CEO of embedded vendor NexEdge Systems, whose latest smart camera uses the Ryzen Edge X800. “Every milliwatt we save translates into a smaller battery, a cheaper housing and, ultimately, a product our customers can deploy at one-tenth the operating cost of a cloud-linked camera.”

Market analysts say the timing is critical. The transformer-optimized AI chip segment is projected to grow 20 % annually through 2034, reaching USD 278 billion, but 40 % of that value will be left on the table unless vendors solve thermal and battery constraints at the edge. “Chipmakers that can deliver 50-plus TOPS inside a 5 W envelope will own the next hardware cycle,” remarked Nina Bhatti, principal analyst at Global Market Insights.

Government policy is adding urgency. The U.S. Department of Energy’s September 2025 efficiency mandate requires all federally funded IoT sensors to operate for five years on a single battery, while the EU’s Ecodesign Directive will cap chipset standby power at 20 mW by 2027. Meeting those targets is pushing foundries toward wide-band-gap materials; STMicroelectronics and Infineon are ramping 200-mm silicon-carbide and gallium-nitride lines that shrink power conversion losses inside edge modules by 35 %.

About NexEdge Systems