Tagged embedded systems · show all articles
Running inference on a battery-powered device instead of a data center is not a smaller version of the same problem. It is bound by a different budget of power, latency, and privacy that reshapes the electronics from the sensor outward.
The lab demo runs once, indoors, tethered to a bench supply. The field unit has to run for years, outdoors, on a battery, without anyone ever opening the enclosure. Nearly every hard problem in edge AI lives in that gap.
A working method for the four decisions that decide whether an edge AI device survives contact with real hardware: how to quantize a model, how to budget its power, how to pick its sensor front end, and where deployments quietly fail.
Three ways to put inference at the edge — dedicated NPU, DSP, and always-on microcontroller — and two ways to sense the world for it — event-based and frame-based — set out as tradeoffs, not a leaderboard.
While most of this cohort tracks gigawatt-scale datacenter deals, SiMa.ai is building AI chips meant to run inside a camera, a robot, or a car — nowhere near a datacenter at all.