EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Bias-Aware FP8 Format for Efficient Weight Representation in Post-Training Quantization

Aleksandra Z. Jovanovic, Stefan S. Tomic, Jelena R. Nikolic, Zoran H. Peric, Davide Bertozzi, Michele Favalli

IEEE Access · 2026 · 14 · 121961-121977

DOI: 10.1109/access.2026.3721067 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Aleksandra Z. Jovanovic und 5 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: European Union’s Horizon Europe Research and Innovation Programme under the Twinning Scheme, U.K. Research and Innovation (UKRI) through U.K. Government’s Horizon Europe, Ministry of Science, Technological Development and Innovation of Serbia.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Deep neural network (DNN) weights dominate memory consumption during inference, making their efficient numerical representation particularly important for resource-constrained Edge AI systems. Widely used 8-bit floating-point (FP8) formats, such as E5M2 and E4M3, provide representable ranges that are significantly wider than required for typical DNN weights, which exhibit low variance and a narrow effective support. Consequently, a substantial portion of the available representation range remains unused. Motivated by the observation that exponent bias scales representable values, we establish a general analytical framework linking FP representation with quantization theory by interpreting FP formats as piecewise-uniform quantizers, enabling systematic analysis of exponent bias effects on representation efficiency. The framework formalizes encoding and decoding rules for low-bit FP representations and provides analytical estimates of representation accuracy through derived expressions for mean squared error and signal-to-quantization-noise ratio (SQNR) as functions of data variance. The analysis shows that the maximum achievable SQNR, primarily determined by mantissa precision, is sustained over a variance range whose width is governed by the exponent–mantissa allocation, while its position is controlled by the exponent bias. This enables increased mantissa precision within a fixed bit budget by compensating for reduced variance coverage caused by fewer exponent bits through an appropriate bias selection that aligns the high-SQNR region with the statistical properties of the data. Based on these insights, we propose Bias-Aware FP8 for Weights (BA-FP8-W). Experimental results on ResNet-18 and DenseNet-121 demonstrate that BA-FP8-W achieves higher SQNR than E4M3 and E5M2 and delivers inference accuracy close to the FP32 baseline.

Förderung

  • European Union’s Horizon Europe Research and Innovation Programme under the Twinning Scheme
  • U.K. Research and Innovation (UKRI) through U.K. Government’s Horizon Europe
  • Ministry of Science, Technological Development and Innovation of Serbia

Zitieren

APA

Jovanovic, A. Z., Tomic, S. S., Nikolic, J. R., Peric, Z. H., Bertozzi, D., & Favalli, M. (2026). Bias-Aware FP8 Format for Efficient Weight Representation in Post-Training Quantization. IEEE Access, 14, 121961-121977. https://doi.org/10.1109/access.2026.3721067

Vancouver

Jovanovic AZ, Tomic SS, Nikolic JR, Peric ZH, Bertozzi D, Favalli M. Bias-Aware FP8 Format for Efficient Weight Representation in Post-Training Quantization. IEEE Access. 2026;14:121961-121977. doi:10.1109/access.2026.3721067

BibTeX
@article{Jovanovic2026,
  author = {Jovanovic, Aleksandra Z. and Tomic, Stefan S. and Nikolic, Jelena R. and Peric, Zoran H. and Bertozzi, Davide and Favalli, Michele},
  title = {Bias-Aware FP8 Format for Efficient Weight Representation in Post-Training Quantization},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {121961--121977},
  doi = {10.1109/access.2026.3721067}
}
Kostenlos starten →