Tiny EdgeAI system‑on‑module

A system-on-module is the part of a device where the hard work lives: processor, memory, and wireless hooks packed onto one tiny board so the rest of the product can be a simpler carrier board. In April 2026, Grinn started pitching a new one, the AstraSOM-261x, as a 25 by 25 millimeter edge artificial intelligence module for devices that need local computing in very little space. (grinn-global.com) “Edge artificial intelligence” just means the model runs inside the device instead of sending every image, sound, or sensor reading to a cloud server. That cuts delay, reduces bandwidth, and keeps more data on the box itself, which is why small cameras, factory sensors, and medical gadgets keep moving in this direction. (grinn-global.com) The surprise is the size. Grinn says the AstraSOM-261x fits a dual-core Arm Cortex-A55 processor, a Mali-G31 graphics unit, and a dedicated neural processor rated at 1 tera operation per second into a 25 by 25 millimeter package with an LGA178 footprint. (grinn-global.com) That footprint matters because a system-on-module is usually the compromise between “design your own computer from scratch” and “buy a whole single-board computer that is too big.” Grinn says all general-purpose input and output pins are exposed, which gives hardware teams room to attach their own sensors, radios, and power circuits without redesigning the compute core. (grinn-global.com) Grinn ties the module to the Synaptics Astra SL2610 family, and its product page lists full high-definition video at 60 frames per second plus “enterprise-grade security” for industrial, medical, and consumer designs. That combination tells you what this class of product is chasing: enough artificial intelligence for vision and inference, but in something small enough to disappear inside the enclosure. (grinn-global.com 1) (grinn-global.com 2) Once hardware gets this small, software architecture starts to matter just as much as silicon. Bare-metal programming means your code runs directly on the chip in a simple main loop with interrupts, while a real-time operating system adds a scheduler so multiple jobs can take turns in a controlled way. (acldigital.com) (rfwireless-world.com) Bare-metal usually wins on overhead because there is no operating system kernel taking memory and processor time. A real-time operating system usually wins when the device has several jobs at once, like running inference, handling networking, logging data, and updating a display without one task blocking the others. (acldigital.com) (codelucky.com) Security is the other half of the story, because tiny connected boxes now get judged on process as much as on features. On April 3, 2026, Microchip said it had earned IEC 62443-4-1 maturity level 2 certification for its industrial automation and control system development process, which is a secure-development standard rather than a claim that every end product is automatically certified. (microchip.com) IEC 62443 is the industrial cybersecurity rulebook many factory and infrastructure buyers now use when they screen suppliers. Microchip’s own explainer says the standard affects how semiconductor parts are selected for industrial internet-of-things products, so a chip vendor with a recognized process can make life easier for device makers chasing regulated customers. (microchip.com 1) (microchip.com 2) Put those pieces together and you get the real engineering choice in 2026. A 25 by 25 millimeter module can solve the space problem, but teams still have to pick the software model that fits their timing needs and the component stack that fits their customer’s security checklist. (grinn-global.com) (acldigital.com) (microchip.com)