Fused-Kernel Hardware Acceleration for Real-Time Stability Monitoring in Kolmogorov-Arnold Networks
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Kolmogorov-Arnold Networks (KANs) represent a major leap in interpretable AI, yet their real-world utility is currently crippled by the ”Memory Wall”—a massive latency bottleneck in spline coefficient evaluation. Standard deep learning frameworks are poorly equipped for this architectural shift, leading to a ”computational tax” that renders KANs impractical for high speed systems. To solve this, we developed the **AS-KAN (Arnold-Stability KAN) Controller**, a hardware-aware fused-kernel specifically optimized for NVIDIA’s Turing architecture.Our approach replaces traditional decoupled stability checks with a real-time Arnold Stability Index (ASI) Governor integrated directly into the Triton-based GPU kernel. By utilizing tiled memory access and shared-memory (SRAM) telemetry, the controller eliminates redundant global memory passes and maintains a deterministic hardware heartbeat during high-throughputinference. Empirical tests on an NVIDIA Tesla T4 GPU show that the AS-KAN Fused-Kernel delivers a consistent **3.1x to 3.5x speedup** over standard PyTorch implementations. Even as workloads scale beyond 120,000 elements, the controller maintains latencies below 0.010ms, effectively slashing the computational overhead of spline networks by over 60%. This work bridges the gap between theoretical spline mathematics and the physical constraints of modern silicon, providing a scalable pathway for explainable AI in edge-deployable hardware.
Fields of Study & Subjects
Parallel Computing and Optimization Techniques
Explainable Artificial Intelligence (XAI)
Advanced Neural Network Applications
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 28 March 2026
- Language
- EN
- Document type
- Preprint