Superconducting neural chip processes fusion data at a million samples per second

A superconducting neural-network chip kept pace with a stream of recorded fusion data arriving at one million samples per second, Brookhaven National Laboratory reported on 19 October. The demonstration tackled a demanding computing problem: recognising rapidly changing plasma conditions quickly enough for the result to be useful in a future control system.[1]

Brookhaven worked with Princeton Plasma Physics Laboratory, the DIII-D National Fusion Facility and chip developer Great Sky. Two tasks used recorded DIII-D measurements for anomaly classification and disruption prediction. A third used TORAX simulations to predict sawtooth instabilities. The demonstration processed those datasets; it did not report closed-loop control of a live plasma.[1]

Why a million samples per second matters

In a tokamak, powerful magnetic fields confine an extremely hot plasma. Its state can evolve rapidly, and some unstable behaviours threaten to interrupt the experiment or damage components. A controller needs fresh measurements, a timely interpretation and an action that arrives before the relevant state has changed. A prediction delivered too late is effectively a historical explanation. Faster computing is valuable only if its output can flow through sensors, decision logic and actuators with suitably low and predictable delay.[1]

Conventional silicon processors are versatile, but researchers are exploring physical alternatives that perform some mathematical operations more directly. Superconducting circuits operate at extremely low temperatures and can support rapid signalling. In the reported system, hardware was designed to implement neural-network-like processing of streams. The point is not that scientists have recreated a human brain inside a fusion machine. The device is a specialised computing substrate for recognition tasks requiring very low latency.[1]

Throughput measures how much data the chip can process. A plasma controller also needs a dependable end-to-end response: measurements must reach the predictor, the prediction must arrive soon enough, and actuators must respond appropriately. The reported million-sample rate addresses the processing step. Prediction accuracy and integration with the full control loop will determine its operational value.[1]

Connecting the cold computer to the hot experiment

There is a physical irony in putting superconducting electronics near fusion research. The plasma must reach enormous temperatures; superconducting circuitry must be kept exceptionally cold. Any future deployment therefore requires infrastructure for thermal isolation, cryogenic operation, reliable interfaces and maintenance. A laboratory chip's computational performance cannot be translated into overall energy efficiency without including those supporting systems. Neither can one assume that a circuit performing well on recorded data will remain stable through the electrical interference and operational interruptions of a real facility.[1]

The TORAX test makes a different contribution. Sawtooth instabilities are recurring changes in the plasma's internal conditions. Demonstrating pattern recognition using simulated data allows researchers to probe specific dynamical behaviours under controlled assumptions. But simulated signals carry the simulator's choices about plasma physics and instrumentation. Testing against real measurements—and eventually under carefully supervised operation—will be essential before treating such a predictor as an element of machine protection.[1]

The result advances a hardware path for plasma analysis rather than a new language-model benchmark. DIII-D’s team says it intends to explore integration during a future experimental campaign. That will bring the key engineering questions together: prediction quality, dependable timing and the interface between a very cold computing device and a very hot physical experiment.[1]

Report an error or suggest a correction