Quantized encoders speed up earth observation on common x86 computers
Quantized Edge Inference of a Self-Supervised SAR-Optical Encoder on Commodity x86 Hardware
Computer Vision and Pattern RecognitionMachine Learning
Summary
Running advanced AI models that analyze Earth images usually needs powerful computers. This paper looks at how a particular AI model, trained to understand satellite images, works on regular x86 computers used in the field. The researchers found that using a simpler number format for calculations (called INT8) speeds up the analysis without losing accuracy and uses less energy. They also show that which part of the computer is fastest changes depending on the computer's capabilities.
What this means in practice
- •For remote sensing engineers: Run Earth observation AI models faster and with less energy on standard x86 hardware by using INT8 quantization without losing accuracy.
- •For edge computing developers: Optimize AI inference speed on low-power CPUs by choosing quantization and hardware configurations informed by these latency and energy insights.
Authors
Christian Sanchez-Leon, Dongming Chen
Abstract
Self-supervised vision transformers are increasingly used as reusable encoders for Earth observation, but their inference cost on the hardware available in the field is rarely characterized. We deploy a pretrained SAR-optical masked-autoencoder encoder on three commodity x86 nodes, from a low-power CPU without AVX2 to an AVX2+VNNI reference, at FP32, FP16 and INT8 on the CPU and, where present, the integrated GPU, and report latency, energy (RAPL and wall plug) and linear-probe accuracy. Quantization is accuracy-neutral: INT8 changes micro-mAP by only +0.0006, within the seed spread, so precision becomes a pure cost decision. INT8 lowers CPU latency on every node, by 1.45x even on the CPU without AVX2, and CPU energy wherever it is measured, and the VNNI reference (1.53x) suggests a further gain beyond AVX2. The Pareto-optimal device shifts from the integrated GPU, which is 4.38x faster than the CPU on the weakest node, to the CPU as the CPU class improves.