RADICAL: Reconfiguration-Aware Direct-Connect AI Cluster using Spatial-and-Wavelength Selective Switching
Abstract
We present RADICAL, the silicon-photonic-enabled Reconfiguration-Aware Direct-Connect AI Cluster that integrates a custom spatial-and-wavelength selective switch (SWSS) with the Linux network stack. RADICAL deploys a packaged 4x4x4λ microring-based SWSS into a four-GPU-server cluster using standard SFP28 optical transceivers, and a reconfiguration-aware host-networking interface to demonstrate a direct-connect reconfigurable optical fabric. The SWSS provides dynamic spatial and wavelength routing with 3.2-6.8 dB path loss, average wavelength and spatial crosstalk isolation exceeding 25 dB, and error-free 25 Gbps signal recovery across routing paths. At the system level, RADICAL validates reconfiguration-aware operation through continuous bulk data transfers totaling 6.4 TBytes over 10 minutes. We further demonstrate multitenant AI training across reconfigurable link topologies, including four ResNet-18 instances and six nanoGPT instances, with the latter sustained for more than five hours.
Related
- Wavelength-selective second-harmonic splitting and focusing in engineered χ ⁽²⁾ photonic crystals
- Orbital-Selective Coexistence of Interlayer Spin-Singlet Formation and SDW Order with Anomalous Spin Reconfiguration in Bilayer Nickelate La3Ni2O7 Revealed by 17O-NMR
- Photoinduced switching of orbital-selective Mott states
- Wavelength-selective bidirectional plasticity in self-powered Se/ZnMgO visual synapses
- Rate-selective synchronization in switching networks: when timing meets structure
BibTeX
@misc{george2026radical,
title = {{RADICAL: Reconfiguration-Aware Direct-Connect AI Cluster using Spatial-and-Wavelength Selective Switching}},
author = {Brett George and Weiyang Wang and Yuyang Wang and Zhenguo Wu and S. C. Mancas and Xiang Meng and Manya Ghobadi and Keren Bergman},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.31941},
eprint = {2610.31941},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.31941},
}
RIS
TY - UNPB
TI - RADICAL: Reconfiguration-Aware Direct-Connect AI Cluster using Spatial-and-Wavelength Selective Switching
AU - Brett George
AU - Weiyang Wang
AU - Yuyang Wang
AU - Zhenguo Wu
AU - S. C. Mancas
AU - Xiang Meng
AU - Manya Ghobadi
AU - Keren Bergman
PY - 2026
DA - 2026/10/22
PB - arXiv
DO - 10.48550/arxiv.2610.31941
UR - https://arxiv.org/abs/2610.31941
AB - We present RADICAL, the silicon-photonic-enabled Reconfiguration-Aware Direct-Connect AI Cluster that integrates a custom spatial-and-wavelength selective switch (SWSS) with the Linux network stack. RADICAL deploys a packaged 4x4x4λ microring-based SWSS into a four-GPU-server cluster using standard SFP28 optical transceivers, and a reconfiguration-aware host-networking interface to demonstrate a direct-connect reconfigurable optical fabric. The SWSS provides dynamic spatial and wavelength routing with 3.2-6.8 dB path loss, average wavelength and spatial crosstalk isolation exceeding 25 dB, and error-free 25 Gbps signal recovery across routing paths. At the system level, RADICAL validates reconfiguration-aware operation through continuous bulk data transfers totaling 6.4 TBytes over 10 minutes. We further demonstrate multitenant AI training across reconfigurable link topologies, including four ResNet-18 instances and six nanoGPT instances, with the latter sustained for more than five hours.
ER -