New edge AI processors from Axeler, Kinara, EdgeCortix, SiMa.ai, and STMicroelectronics enable local inference

Recent advancements in edge AI hardware are enabling complex generative and vision-based inference tasks to run locally on low-power devices, reducing reliance on cloud-based infrastructure. Several companies have introduced specialized processors designed to handle these workloads efficiently. Axeler’s Metis AI-PU utilizes a proprietary digital in-memory compute architecture to increase throughput and reduce data movement. Kinara’s Ara-2 processor, recently acquired by NXP, supports generative AI models like Llama2-7B and Stable Diffusion 1.4 in embedded environments. EdgeCortix’s Sakura-II features a runtime reconfigurable fabric, while SiMa.ai’s MLSoC focuses on software-led, edge-native deployment for video workloads. Finally, STMicroelectronics has integrated its Neural-ART accelerator into the STM32N6 microcontroller, bringing machine learning capabilities to a widely deployed MCU family. These developments demonstrate that real-time generative AI is increasingly viable at the low-power edge, offering improved privacy, lower latency, and reduced network dependency for applications ranging from robotics and industrial automation to smart retail and consumer devices.

Axeler’s Metis AI-PU uses a proprietary digital in-memory compute architecture where every SRAM cell doubles as a compute unit. Kinara’s Ara-2 processor supports generative AI models and can run multiple models concurrently on the same input stream.

EdgeCortix’s Sakura-II utilizes a runtime reconfigurable fabric to dynamically optimize compute unit interconnections for each specific model. SiMa.ai’s MLSoC is a software-led, edge-native system designed to handle multiple 4K video streams simultaneously under 10 watts.

STMicroelectronics has integrated its Neural-ART accelerator into the STM32N6 microcontroller to bring machine learning to a widely used MCU family. These new edge AI processors aim to reduce latency and improve privacy by enabling inference to occur closer to the data source.