EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

QTN-Hybrid: A DSP-Free Non-Linear Tensor-Network Classifier for Low-Latency FPGA Inference

Kyung-Won Cho, Sung-Hyun Go, Sungkwan Youm

IEEE Access · 2026 · 14 · 163350-163368

DOI: 10.1109/access.2026.3745871 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Kyung-Won Cho und 2 weiteren Autor:innen. Die Arbeit nennt 21 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Wonkwang University in 2026.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Tensor-network (TN) classifiers based on matrix-product states (MPS) have recently emerged as a route to fully DSP-free inference on FPGAs, but to the best of our knowledge, every reported deployment has been strictly multilinear. We address this gap by introducing QTN-Hybrid, a quantum-inspired tensor-network (QTN) hybrid design in which a DSP-free non-linear MPS classifier inserts a LUT-fabric-mapped Hardtanh activation (implemented as two signed comparators and a three-way selection multiplexer) between bond contractions. The Hardtanh non-linearity bounds the hidden vector after every bond contraction and thereby confines the fixed-point dynamic range along the chain, so that accuracy is retained at low precision while hardware resources are reduced at the same time. The proposed model is trained on Fashion-MNIST at an operating point comparable in size to the published cascaded spaced-matrix-product-operator (CSMPO) reference design ( $\chi =4$ , thirty-two sites, $\approx 1{\,}$ k parameters), repeated over five random seeds, and evaluated after post-training quantization at precisions from 4–16 bits. At every precision QTN-Hybrid is more accurate than the linear MPS, by + 4.38–4.50 pp at 8–16 bits and by + 5.30 pp at 6 bits, and at 6 bits it remains above the 16-bit linear baseline. On the hardware side, the kernels are C-synthesised by high-level synthesis at a $250{\,}$ MHz target clock and the generated RTL is logic-synthesised to obtain resource and timing figures. The resulting designs use zero DSP blocks and require several times fewer LUTs than the linear 16-bit baseline, and both the 8-bit and the 6-bit configurations pass post-synthesis resource-capacity and timing checks on a mid-range FPGA and on a low-cost edge system-on-module, indicating at the post-synthesis level that deployment on small FPGAs is feasible, while routability and on-board operation remain to be verified.

Förderung

  • Wonkwang University in 2026

Zitieren

APA

Cho, K., Go, S., & Youm, S. (2026). QTN-Hybrid: A DSP-Free Non-Linear Tensor-Network Classifier for Low-Latency FPGA Inference. IEEE Access, 14, 163350-163368. https://doi.org/10.1109/access.2026.3745871

Vancouver

Cho K, Go S, Youm S. QTN-Hybrid: A DSP-Free Non-Linear Tensor-Network Classifier for Low-Latency FPGA Inference. IEEE Access. 2026;14:163350-163368. doi:10.1109/access.2026.3745871

BibTeX
@article{Cho2026,
  author = {Cho, Kyung-Won and Go, Sung-Hyun and Youm, Sungkwan},
  title = {QTN-Hybrid: A DSP-Free Non-Linear Tensor-Network Classifier for Low-Latency FPGA Inference},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {163350--163368},
  doi = {10.1109/access.2026.3745871}
}
Kostenlos starten →