Insider Brief
- Great Sky and research partners demonstrated a superconducting AI chip on fusion-related prediction tasks, supporting its potential use in real-time plasma control.
- The chip processed DIII-D tokamak data at one million samples per second for anomaly classification and disruption prediction, and used simulated data to predict sawtooth instabilities.
- The chip implements analog neural networks using superconducting circuitry and optical communication, with a next-generation chip featuring a neural network 10 times larger planned for demonstration in December.
PRESS RELEASE — Great Sky, Brookhaven National Lab, Princeton Plasma Physics Laboratory, and the DIII-D National Fusion Facility partnered to develop a new kind of AI chip with the speed necessary to sense and respond to searing plasmas present in fusion reactors. The chips differ from conventional silicon chips — such as CPUs and GPUs — in fundamental ways. Instead of processing digital information using silicon transistors, Great Sky’s chips implement analog neural networks using superconducting circuitry with optical communication.
The team designed and fabricated their latest superconducting neural-network chip and demonstrated the chip performing three tasks of relevance to fusion control. The tasks were selected due to their relevance to the operation of tokamaks — large reactors designed for producing fusion power.
Two of the tasks injected data collected at the DIII-D tokamak into the chip for anomaly classification and disruption prediction. The data was collected with a very high sample rate of a million samples per second, and the team showed the chip had no problem keeping up with that speed.
The third task processed simulated data generated by TORAX — a tokamak simulator developed by the Google DeepMind Fusion team. For this task, the chip was able to process simulated signals to predict sawtooth instabilities, an event that is critical to tokamak operation. This example illustrates the feasibility of using Great Sky’s chips as part of the real-time simulation of plasma state that is central to tokamak control and stabilization during operation.
“It’s incredible to see the team processing data from our tokamak,” said Richard Buttery, Director of DIII-D. “The folks here at DIII-D are very excited to see how this chip can impact diagnostics and control for the fusion industry.”
Superconducting neural networks sit at the intersection of several DOE priorities: advanced AI, quantum technologies, and transformative energy. This collaboration shows how expertise in scientific AI and advanced computing can help move promising hardware concepts toward real-time control challenges in fusion and other complex systems.
Fusion reactors demand control systems that can think at the speed of a rapidly evolving plasma. This demonstration showed a chip taking data straight from a tokamak at a million samples per second without missing a beat. It is a meaningful step toward the real-time control systems that practical fusion power will require.
The Great Sky team has been developing this technology for over a decade, first at the National Institute of Standards and Technology before spinning out to form Great Sky. Their approach to AI hardware brings significant advantages in speed and energy efficiency, directly addressing the problems that hinder the economic foundations of the AI and data-center industries. This demonstration represents a significant step forward for the field of superconducting neural networks.
“This is our largest chip to date, and it’s great to see it processing high-speed data directly measured from a tokamak,” said Jeff Shainline, founder and CEO of Great Sky. “But this is just the beginning. This approach to AI hardware has so much to offer with regard to fusion, as well as other applications at the forefront of technology. Control and error correction of quantum systems, advanced cybersecurity, video intelligence with continual learning — these are all incredible opportunities as we scale up our systems.”
The team will demonstrate their next-generation chip with a 10x larger neural network in December.
