The datacentre is the land of giant GPUs. The edge is a zoo. Hearing aids, cameras and sensors cannot spend watts on matrix multiplies, so vendors are reviving exotic architectures that trade precision for power. Several are now in consumer devices.
Precision is negotiable
Most edge workloads tolerate noisy arithmetic. Analog multiply-accumulate arrays exploit that, delivering useful accuracy at microwatt budgets. The catch is tooling: every unusual chip needs its own compiler, and that is where most startups stumble.
Key takeaways
- Analog and in-memory designs are in production devices
- Microwatt inference enables always-on sensing
- Compilers remain the main bottleneck
This story is part of our ongoing coverage of ai. We will update it as more details emerge.
