HWBRIEF
Nottingham Trent, Imperial and Aston researchers say the peak trillions of operations per joule number on chip datasheets — TOPS/W — measures the chip, not the system, as AI moves into phones, wearables, and battery powered sensors.
The headline efficiency number on every AI chip datasheet (peak TOPS/W, the trillions of operations a chip can claim per joule in its best moment) measures the chip, not the system. A new framework paper from Nottingham Trent University, Imperial College London, and Aston University argues that is the wrong denominator as AI compute moves from data centers to phones, wearables, drones, and battery-powered sensors.
The authors, Eiman Kanjo and Varuna De Silva, want a system-level lens. A digital controller would decide when an analogue, photonic, or neuromorphic substrate (silicon that computes inside memory, with light, or with brain-like spikes) actually earns its place, and how to measure whether it did. The framework covers four paradigms on one axis: digital, analogue in-memory, photonic, and neuromorphic.
The paper also surveys the digital-only levers that still work: pruning, quantisation, knowledge distillation, sparsity, and event-driven sensing on small microcontrollers with hundreds of kilobytes of SRAM. Those controllers can already run ImageNet-scale inference, the authors note, which is part of why the metric matters more now.
No vendor benchmark or independent deployment data has been published alongside the preprint, posted to arXiv on 4 August 2026 and not yet peer reviewed. The trigger coverage summarises the same argument. The authors' framing: physical substrates earn a role when they improve the system, not when they top a datasheet.