Sustained-load thermal and activation-memory characterization

Characterize the sustained-load thermal behavior and peak activation memory of AgroVisNet on the intended embedded deployment hardware, including comparison with the evaluated baseline models.

Background

The paper measures parameter count, multiply–accumulate cost, model size, quantization behavior, desktop CPU latency, and latency and throughput on a mid-range Android handset. However, those measurements do not establish how the quantized model behaves under sustained inference workloads or how much peak activation memory it requires.

The authors explicitly state that sustained-load thermal behavior and peak activation memory remain unresolved, leaving an important part of the deployment claim unverified for low-cost embedded hardware.

References

The efficiency comparison now covers multiply-accumulate cost, single-precision and quantized model size, desktop CPU inference latency and, in Table~\ref{tab:edge}, on-device latency and throughput on a mid-range Android handset, but sustained-load thermal behaviour and peak activation memory remain open, the latter reported neither on-device nor for the baselines.

— AgroVisNet: A lightweight Convolutional Network and the BD-PlantDX Expert-Validated Benchmark for Radish, Potato and Pointed Gourd Disease Classification  (2609.10469 - Mandal et al., 9 Sep 2026) in Section 7, Discussion and Limitations; Section 8, Conclusion