Real-Time Intelligence for Physical AI at the Edge

By Durgesh Srivastava, MIPS CTO

March 5, 2025

The age of physical AI is approaching, taking the virtual world of digital twins and move it into the physical world of autonomous platforms that can understand the world around them and make decisions, and quickly adapt to new tasks. Today’s robots specialize in narrowly defined tasks, built for simple automation of repetitive behavior. The next generation of autonomous platforms will be more advanced and adaptable to new tasks, environments, and events.

To address this opportunity, the MIPS S8200 physical AI subsystem is designed to enable a wide range of physical AI applications at the edge, or endpoint. Physical AI is a combination of safe, high-reliability data processing combined with mixture-of-experts AI model support. Foundational models for action, vision, and language are combined with reinforcement learning to expand the use of advanced, precise physical expertise in many industries.

In the MIPS Atlas portfolio, the MIPS S8200 physical AI subsystem enables a broad range of applications to be more adaptable. Autonomous robots will augment our factories, workspaces, offices, and transportation systems. Growing and harvesting food may be made more efficient, manageable, and consistent for better yield per acre. Medical platforms can leverage the techniques of predictive maintenance to alert health care professionals to new changes in a patient. The benefits of autonomous platforms capable of acting without human supervision in a specific role or function continue to grow.

The MIPS S8200 will feature highly efficient RISC-V application cores with tightly coupled AI engines. Capable of running vector or matrix workloads, these engines provide support for post-generative AI foundation models and enable a mixture-of-experts approach. Physical AI models with combinations of action, visual, and language models will be used to perform tasks without retraining to embody intelligence and adaptability.

Aiming to enable a 4X higher TOPS/W (tera-operations per second, per Watt) output, the MIPS S8200 is built out of quad-core building blocks, featuring multiple-thread support, and tiling of coherent clusters for scaling up into large designs, offering performance ranges from 10 TOPS to 1000’s of TOPS. With industry-leading performance combined with a power-efficient, open-specification based design, adopting physical AI in autonomous platforms will be easier than ever before.

MIPS is an exhibitor at Embedded World 2025 from March 11^th – 13^th, in Nuremburg, Germany. To request a meeting, please contact us here: MIPs.com/events