Metraipolis › archive › dc_66da978f2866

A Time-Multiplexed Q16.16 Fixed-Point Neural Network Accelerator for Resource-Efficient FPGA Inference (2026) · DOI 10.20944/preprints202608.2059.v1

Source
Metraipolis archive
Original
https://doi.org/10.20944/preprints202608.2059.v1
License (as archived)
CC-BY-4.0
Archived
2026-10-08T01:11:59.961Z
SHA-256 of the archived text
ba7faea9c364ae022dd64524daa748a680c8f22120d81f3721e8250fb2130cbb
Stable URL of this page
https://metraipolis.com/documento/dc_66da978f2866

How to cite

A Time-Multiplexed Q16.16 Fixed-Point Neural Network Accelerator for Resource-Efficient FPGA Inference (2026) · DOI 10.20944/preprints202608.2059.v1 — Metraipolis archive, https://doi.org/10.20944/preprints202608.2059.v1 (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_66da978f2866 (Metraipolis archive dc_66da978f2866, sha256 ba7faea9c364ae022dd64524daa748a680c8f22120d81f3721e8250fb2130cbb)
BibTeX
@misc{dc66da978f2866,
  title = {A Time-Multiplexed Q16.16 Fixed-Point Neural Network Accelerator for Resource-Efficient FPGA Inference (2026) · DOI 10.20944/preprints202608.2059.v1},
  howpublished = {Metraipolis archive},
  url = {https://doi.org/10.20944/preprints202608.2059.v1},
  note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_66da978f2866; sha256 ba7faea9c364ae022dd64524daa748a680c8f22120d81f3721e8250fb2130cbb},
  year = {2026}
}

Archived text

Reproduced under CC-BY-4.0. Attribution: Metraipolis archive, https://doi.org/10.20944/preprints202608.2059.v1. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Articolo recente con licenza aperta: scheda dai metadati Crossref (posted-content).
Titolo: A Time-Multiplexed Q16.16 Fixed-Point Neural Network Accelerator for Resource-Efficient FPGA Inference
Autori: Babak Alivand Bonab, Soheila Gharavi Hamedani
Pubblicato: 2026-08-28 · editore: MDPI AG
DOI: https://doi.org/10.20944/preprints202608.2059.v1
Licenza dell'articolo: CC-BY-4.0

Abstract (testo degli autori, CC-BY-4.0):
Deploying neural networks on edge devices requires a careful balance of speed, power, and area. FPGAs are well-suited for this task, but conventional floating-point arithmetic can incur significant hardware overhead. To address this challenge, this paper presents a multi-layer feedforward neural network accelerator designed for FPGA-based inference. The architecture employs a custom Q16.16 fixed-point format, offering a practical balance of numerical precision and hardware efficiency compared to floating-point units. A time-multiplexed data-path with a 64-bit accumulator, processes computations sequentially using a single MAC unit per neuron, aiming to reduce hardware consumption. This resource-conscious structure allows the network size to scale with modest increases in logic utilization. Deterministic saturation logic is implemented to help prevent overflow and contribute to stable operation under varying numerical conditions. The design was implemented entirely in VHDL and synthesized for an Artix UltraScale+ FPGA, achieving a maximum operating frequency of 139 MHz for a representative 4‐8‐8‐4 network, the accelerator requires only 520 LUTs, 1182 flip‐flops, and 4 DSP slices, corresponding to less than 1% of the available logic resources and requiring zero BRAM. One inference is completed in 233 clock cycles, yielding a throughput of approximately 597 K inferences per second. Verification using unit‐level testbenches and a 10,000‐iteration randomized stress test achieved 100% functional stability.

Fonte: https://api.crossref.org/works/10.20944%2Fpreprints202608.2059.v1 — metadati Crossref, consultati il 2026-10-07. Interrogazione: «efficient inference quantization».