ResearchResearch paperEfficiency & Inference · Training & Scaling · Hardware & Compute1 source · Oct 12, 2026

LocalTrain: Integer-Only Local Learning for Energy-Efficient CNN Training on Edge FPGAs

We present an FPGA system that trains a CNN and its regional auxiliary classifiers using a shared integer matrix engine.

Local learning reduces cross-layer training dependencies, but an efficient accelerator must still retain operands, accumulate gradients, and select quantization scales.
Compact errors are expanded on chip, gradients are folded tile by tile, and historical scales permit direct output conversion.
Effective matrix throughput is 31.22 and 31.57 GMAC/s, respectively.
Against the same integer method on Jetson Orin Nano, measured training energy is 40.4% and 42.1% lower, with $1.86\times$ and $1.82\times$ longer runtime; measurements use the respective platform power domains.

[1]LocalTrain: Integer-Only Local Learning for Energy-Efficient CNN Training on Edge FPGAs
arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 12, 04:57 PM
We present an FPGA system that trains a CNN and its regional auxiliary classifiers using a shared integer matrix engine.
Local learning reduces cross-layer training dependencies, but an efficient accelerator must still retain operands, accumulate gradients, and select quantization scales.

Extractive summary: sentences quoted from the sources.

Before this