Real-Time Inference for Control
Running a model fast and reliably enough to sit inside a plasma control loop.
The latency budget
A plasma control loop runs at kilohertz rates, leaving microseconds to tens of microseconds for any model that feeds it. A model that is accurate but slow is useless in the loop. Real-time inference is the engineering of making a trained model produce outputs within a hard, guaranteed time bound.
What must be guaranteed
- Bounded worst-case latency, not just average latency
- Deterministic behavior, no unpredictable pauses
- Graceful handling of missing or late inputs
- Fixed, predictable memory and compute use
Making models fast
Techniques include using compact architectures, quantizing weights to lower precision, pruning unneeded connections, and compiling the model to the target hardware. Fixed-topology networks give constant-time inference, which is why they are favored over data-dependent computation in the loop.
Hardware
Control-loop inference runs on FPGAs, GPUs, or dedicated processors chosen for deterministic timing. The model, its inputs, and its outputs are integrated into the real-time control system with the same rigor as any other safety-relevant component, including verification of timing under worst-case load.
Reliability over cleverness
In the loop, predictable and robust beats marginally more accurate. A model that occasionally stalls or behaves oddly on an unfamiliar input is worse than a simpler one that always responds in time and defers safely when unsure. For design-stage devices such as the Kronos machines, these requirements shape control studies now, so that any future implementation inherits a design built for real-time guarantees rather than retrofitted for them.