Compute at the electrode.
Modern invasive BCIs stream raw neural data off the implant and process it outside the body. That architecture worked at a hundred channels. It breaks at ten thousand. Thalamic moves the processing onto the implant itself.
Channel counts are growing exponentially. Everything downstream is not.
Every electrode adds another stream of wideband data. Moving it all off the implant costs bandwidth the link doesn't have and power the tissue can't absorb — the thermal budget of neural tissue is measured in single-digit milliwatts. The result: systems that record more than they can use, throttled at the interconnect.
Every stage that can run on the implant, runs on the implant.
The T-1 takes the wideband signal where it is generated and reduces it in place. Filtering, spike detection, feature extraction and compression run as a fixed-function chain; decoding runs on top of it. What crosses the skull is never the raw waveform.
How far down that chain a system stops is a design decision, and not one we think should be fixed in silicon. The same datapath can hand back detected events, a compressed signal, or a decoded estimate — the right level depends on the application, the link budget and how much of the task belongs on the implant at all.
Every microwatt becomes heat in tissue.
Power is not a performance metric here, it is a safety constraint: the implant shares its thermal budget with the cortex around it. That rules out running a general-purpose processor at the electrode, and it rules out moving the raw signal out to one.
The architecture is event-driven end to end. Between spikes there is nothing to compute, and the chain spends nothing computing it — energy is drawn when neurons fire, in proportion to how many fire. Activity, not channel count, sets the power draw.
Figures on this page are design targets for a pre-silicon architecture, not measured results.
A processing unit per module, not per system.
The T-1 is designed as a tile. One die serves a module of channels; a system is built by placing as many as the array needs, each reducing its own share of the signal locally. Channel count grows by adding dies, so the interconnect off the implant does not have to grow with it.
Hundreds to low thousands of channels, served by a single processing tile.
Ten thousand channels and above — tiles placed per module, telemetry unchanged.
The regime the architecture is built for: reduction stays local, and only the reduced output is aggregated.