This week, we pushed the boundaries of AI-driven networking, edge storage, and in-network AI, proving that innovation is all about smarter, faster, and more efficient systems. Here’s what made headlines:
🤖 AI/ML for 6G & B5G Networks
- Predictive retraining replaces static schedules: Using LSTM and Local Outlier Factor (LOF), we detect new traffic patterns in real time, cutting SLA violations by 40% and eliminating unnecessary retraining.
- Threshold-based retraining fails—our adaptive approach ensures AI/ML models stay accurate and efficient in dynamic 6G environments.
- Edge vs. Cloud retraining? We balance latency and computational power, optimizing resource use for B5G networks.
💾 Constraint-Aware Edge Storage
- CATER framework integrates with Apache Ozone, reducing active nodes by 23% while respecting GDPR, hardware specs, and collocation rules.
- Heuristics + Optimization: Combining ILP and heuristics, CATER cuts computation time by 90%—delivering real-time adaptability with zero constraint violations.
🌐 In-Network AI & P4 Switches
- DNNs in P4 switches? Our LUT distillation maps complex models into lookup tables, enabling wirespeed DDoS mitigation with 93% F1-score—no ALUs required!
- Lossless accuracy: Traditional quantization sacrifices performance; our method deploys DNNs in P4 hardware without information loss.
💡 GVirtuS + RDMA: HPC Revolution
- Pre-registered buffers supercharge RDMA, slashing context switches by 98% and boosting GPGPU virtualization speed by 55%.
- Matrix Multiplication 5.5x faster—proving Infiniband is the future of low-latency HPC.
🔗 Explore our research: https://www.cleverproject.eu/
📄 Read full papers: