On-device inference
Neural inference that stays on the device.
We're designing an accelerator for running transformer and convolutional workloads at the edge — no host offload, no network round-trip.
This page is being rebuilt. Detailed architecture notes and the performance targets we're designing toward are in preparation — and will be published together with the methodology used to measure them, not before. If you want the technical write-up when it's ready, get in touch.