A superconducting neural-network chip has processed fusion-experiment recordings at their original rate of one million samples a second. Brookhaven National Laboratory, Princeton Plasma Physics Laboratory, the DIII-D facility and Great Sky reported the demonstration on 19 October. The prototype tackles a demanding computing problem: extracting useful signals quickly enough to help an instrument respond to rapidly changing plasma.[1]
Three plasma tasks on specialised hardware
The first two tasks used recorded DIII-D signals: classification of anomalous patterns and prediction of disruptions. The chip kept pace with replayed data sampled at a million measurements per second. A third task used TORAX-generated signals to forecast sawtooth oscillations. These were hardware demonstrations on recorded and simulated inputs, with live control of a tokamak still ahead.[1]
Throughput is one component of a control system. The useful interval runs from a physical change through sensor acquisition and inference to an actuator’s response. A fast classifier can shorten part of that interval; integration with the instrument determines whether the resulting warning arrives early enough and can drive an effective action. That makes the planned DIII-D deployment an important transition from a processing test to work at the machine.[1]
The prototype’s three tasks
| Test | Data entering the chip | What was shown |
|---|---|---|
| Anomaly classification | Previously recorded DIII-D data | Selected plasma patterns classified at high data rate |
| Disruption prediction | Previously recorded DIII-D data | A prediction task processed on the device |
| Sawtooth-instability forecasting | TORAX simulated signals | Relevant simulated plasma behavior tracked |
Why superconducting hardware is interesting
The device implements an analogue neural network in superconducting circuits, with optical communication connecting its components. Superconducting operation offers a route to fast processing at very low electrical resistance, while the complete system still includes cooling and peripheral equipment. Its immediate attraction here is the timing of inference close to a scientific instrument, where transferring data to a distant computer can consume part of the available response window.[1]
TORAX gives the team a way to generate plasma behaviour for the sawtooth task without waiting for another experimental discharge. Recorded DIII-D data supply a complementary test grounded in an operating tokamak. Using both lets the collaboration exercise the chip on different kinds of relevant signals as it prepares the next hardware generation.[1]
A larger network headed for DIII-D
Brookhaven says a scientific paper is in preparation. The partners expect to demonstrate a chip with a ten-times-larger neural network in early 2027 and deploy it at DIII-D later that year. Those are the project’s announced milestones. The published demonstration so far establishes that the prototype can run the three selected tasks, including high-rate replay from the tokamak.[1]
The larger chip will give the team more room to model plasma signals, while work at DIII-D will put the computing hardware alongside the instrument it is meant to serve. The engineering opportunity is specific: move the calculation close enough to the measurement that rapid physical changes can be acted on while they still matter.[1]
Sources & further reading
- First-of-a-Kind Super Intelligence Chips Successfully Analyze Fusion Plasmas Brookhaven National Laboratory · 19 Oct 2026 · Accessed 25 Oct 2026