Anker is entering the silicon design arena with a processor engineered for local AI in compact devices. Named Thus, the chip employs a compute-in-memory architecture, a significant departure from conventional designs that separate model storage from computation. This approach, where calculations happen directly within the memory storing the neural network parameters, reduces data movement. The result is a chip that claims higher energy efficiency and the ability to run larger models within the strict thermal and spatial budgets of wearables.

The initial deployment target is telling: premium earbuds. This is a deliberate stress test for the architecture. Earbuds present a severe engineering constraint—minimal space for a battery and a need for always-on processing. Previous solutions were limited to small neural networks of a few hundred thousand parameters. Anker states Thus can manage networks with several million parameters, enabling more complex audio processing tasks like advanced environmental noise cancellation.

System design here extends beyond the chip. Anker's planned earbuds pair the Thus processor with eight MEMS microphones and two bone conduction sensors, creating a data pipeline intended to isolate speech with higher fidelity. The claimed improvement hinges on the chip's ability to apply a more sophisticated model to this richer sensor data in real time.

The trade-off, as with any novel architecture, is unproven real-world performance against established competitors like Apple and Sony. If the efficiency gains materialize, they could validate compute-in-memory as a viable path for on-device AI in power-constrained environments. Full technical specifications and benchmark results are expected at Anker's upcoming product event.

Source: The Verge