Heidelberg University researchers report that heating BrainScaleS-2, a processor that uses electrical signals to calculate neural-network outputs, introduces systematic errors that noise-resistant training only partly addresses. Their September 14 paper points to temperature-aware calibration and training on physical hardware as remedies.
The hardware's design explains why temperature matters. As an earlier system paper describes, BrainScaleS-2 can turn a neural network's multiplication and addition into an electrical operation. Input numbers determine how long synapses, the circuit elements holding connections, remain active. Their currents accumulate in neuron circuits, and converters turn the resulting voltages back into numbers for the next processing step. Computation therefore depends on the physical response of these circuits. A change in that response changes the numbers passed onward, even if the stored network weights and input remain the same.
Engineers already have a way to adapt a network to such hardware imperfections. The earlier BrainScaleS-2 work describes hardware-in-the-loop training: the chip calculates a network's outputs, while a host computer uses those outputs to update its weights. This gives learning access to the device's actual behavior. Software also divides a larger network into operations the chip can execute. For developers, that separation means the model, its execution software and the particular chip must be considered together when diagnosing a wrong output.
The new experiment tested how that behavior changes with heat. A heat gun raised one device from about 40 to 90 degrees Celsius before cooling. Measurements separated random fluctuations from persistent circuit-specific deviations. The authors found that systematic effects dominated the worsening signal error. They then compared remedies on a three-layer classifier recognizing ten speech-command classes.
Adding noise during software training improved robustness against a baseline trained without it; varying the noise strength added a modest benefit away from the calibration temperature. Hardware-in-the-loop fine-tuning generally performed best. Calibration settings were prepared at 40, 60 and 80 degrees Celsius, but hardware training took place at 40 degrees. The best operating behavior depended on both the training calibration and the calibration used during inference.
Those results cover one chip and a small speech task; they do not establish reliability across production devices or large language models. For an engineer considering this approach, the next useful comparison would combine accuracy across the intended temperature range with the time needed to calibrate and retrain each device. Repeating that comparison across chips would show whether a single preparation procedure transfers or whether individual units need their own adjustments. Such measurements are needed before translating this laboratory remedy into a deployment plan.
Sources
- Niklas Summ and colleagues, Beyond Noise: Understanding and Overcoming Temperature Effects in Analog DNN Inference. September 14, 2026; arXiv:2609.15527v1. Temperature protocol, speech task and comparison of remedies.
- Christian Pehle and colleagues, The BrainScaleS-2 Accelerated Neuromorphic System With Hybrid Plasticity. February 24, 2022. Earlier system design and hardware-in-the-loop method; its results are not the new temperature experiment.
- Photograph: Universität Heidelberg, Human Brain Project media library, HICANN-X chip. Reused from the original host without editing the photograph, under CC BY-ND 4.0. This is a photograph of the hardware platform, not a record of the reported heating experiment.