<p>Sixth-generation (6 G) wireless networks demand cognitive radio systems that simultaneously achieve sub-millisecond latency and sustainable energy consumption – requirements conventional artificial intelligence approaches cannot meet. This paper presents a hardware-software co-design framework integrating neuromorphic computing with cognitive radio to address both constraints through brain-inspired spiking neural networks (SNNs). We systematically analyze five neuromorphic platforms – Intel Loihi 2, IBM TrueNorth, SpiNNaker, SpiNNaker 2, and Intel Hala Point – using standardized benchmarks from the Intel Neuromorphic Deep Noise Suppression (N-DNS) Challenge, demonstrating sub-millisecond spectrum decisions (50-170 <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\mu \)</EquationSource> </InlineEquation>s end-to-end latency) with energy consumption reduced by 100-1000<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> (31 pJ per spike) compared to conventional GPU-based approaches (2.5−12.5 <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\mu \)</EquationSource> </InlineEquation>J per operation). Our framework provides three novel contributions: (1) a unified co-design methodology optimizing spike encoding, network topology, and hardware mapping jointly to achieve 3<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\times \)</EquationSource> </InlineEquation> efficiency gains over independent optimization; (2) quantitative design rules for encoding selection – rate coding for signal-to-noise ratios below -10 dB, temporal coding for latency requirements below 100 <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(\mu \)</EquationSource> </InlineEquation>s, and population coding for reliability exceeding 99.9%; and (3) experimental validation achieving 97.6% classification accuracy on real-world spectrum data from industrial IoT deployments consuming only 31 mW average power. Through five detailed case studies spanning industrial automation (99.9% uptime over 6 months), vehicle-to-everything communications (98.7% collision avoidance), defense applications (95% reliability under 40 dB jamming), smart cities (100,000 sensors), and healthcare (15-year implant lifetime), we demonstrate neuromorphic cognitive radio’s practical viability. The framework addresses critical deployment barriers including device variability mitigation (±20% threshold compensation), cross-platform algorithm portability, and RF-to-spike conversion interfaces. These results establish neuromorphic computing as a foundational technology for energy-constrained, latency-critical 6 G wireless systems, with implications extending to radar processing, electronic warfare, and satellite communications.</p>

Energy-efficient neuromorphic computing for ultra-low latency cognitive radio: a hardware-software co-design framework for 6 G spectrum intelligence

期刊论文
出版模式：

开放获取
发表日期：

2026年3月17日

查看全文

Serhiy O. Semerikov,
Pavlo P. Nechypurenko,
Tetiana A. Vakaliuk,
Iryna S. Mintii,
Andrii O. Kolhatin

摘要

Sixth-generation (6 G) wireless networks demand cognitive radio systems that simultaneously achieve sub-millisecond latency and sustainable energy consumption – requirements conventional artificial intelligence approaches cannot meet. This paper presents a hardware-software co-design framework integrating neuromorphic computing with cognitive radio to address both constraints through brain-inspired spiking neural networks (SNNs). We systematically analyze five neuromorphic platforms – Intel Loihi 2, IBM TrueNorth, SpiNNaker, SpiNNaker 2, and Intel Hala Point – using standardized benchmarks from the Intel Neuromorphic Deep Noise Suppression (N-DNS) Challenge, demonstrating sub-millisecond spectrum decisions (50-170 \(\mu \) s end-to-end latency) with energy consumption reduced by 100-1000 \(\times \) (31 pJ per spike) compared to conventional GPU-based approaches (2.5−12.5 \(\mu \) J per operation). Our framework provides three novel contributions: (1) a unified co-design methodology optimizing spike encoding, network topology, and hardware mapping jointly to achieve 3 \(\times \) efficiency gains over independent optimization; (2) quantitative design rules for encoding selection – rate coding for signal-to-noise ratios below -10 dB, temporal coding for latency requirements below 100 \(\mu \) s, and population coding for reliability exceeding 99.9%; and (3) experimental validation achieving 97.6% classification accuracy on real-world spectrum data from industrial IoT deployments consuming only 31 mW average power. Through five detailed case studies spanning industrial automation (99.9% uptime over 6 months), vehicle-to-everything communications (98.7% collision avoidance), defense applications (95% reliability under 40 dB jamming), smart cities (100,000 sensors), and healthcare (15-year implant lifetime), we demonstrate neuromorphic cognitive radio’s practical viability. The framework addresses critical deployment barriers including device variability mitigation (±20% threshold compensation), cross-platform algorithm portability, and RF-to-spike conversion interfaces. These results establish neuromorphic computing as a foundational technology for energy-constrained, latency-critical 6 G wireless systems, with implications extending to radar processing, electronic warfare, and satellite communications.