Towards implementation of neural networks for non-coherent detection MIMO systems - CEA - Commissariat à l’énergie atomique et aux énergies alternatives Access content directly
Conference Papers Year : 2022

Towards implementation of neural networks for non-coherent detection MIMO systems

Abstract

In this paper, we propose the use of quantized neural networks to perform non coherent MIMO detector in sub-TeraHertz (THz) communications. Implementing neural networks is challenging because operations are performed using a high number of bits. This results in slow and energy consuming computations. Then, quantization appears to be essential to consider low latency and energy efficient communication systems. Specifically, in this work, we propose quantizing our designed Neural Network (NN) performing demapping operation. We employ VSORA's digital signal processor (DSP) architecture to perform the quantization. We observe the impact of quantization on the bit error rate. Moreover, we also evaluate the power computation of the proposed DSP regarding our NN. Our simulation results show that we can quantize the weights of the NN to only 6bits with neglectable degradation on the performance. Besides, we expect achieving high throughput (> 1Gbps), with a peak power consumption of only 0.58W. Thus, the proposed quantization scheme and DSP design allow to achieve high throughput and high energy efficiency.
Fichier principal
Vignette du fichier
307-50328-review.pdf (350.33 Ko) Télécharger le fichier
Origin Publication funded by an institution

Dates and versions

cea-03856556 , version 1 (16-11-2022)

Identifiers

  • HAL Id : cea-03856556 , version 1

Cite

Alexis Falempin, Julien Schmitt, Trung Dung Nguyen, Jean-Baptiste Doré. Towards implementation of neural networks for non-coherent detection MIMO systems. VTC 2022 - IEEE 96th Vehicular Technology Conference, Sep 2022, London, United Kingdom. ⟨cea-03856556⟩
38 View
74 Download

Share

Gmail Mastodon Facebook X LinkedIn More