Deploying Physics Models at the Edge: Quantized AI for Weather & Fluid Simulation
summary
Claim: neural PDE solvers replace iterative numerical methods and run on a palm‑sized Dell Pro Max after pruning and quantization, delivering about one petaflop throughput on FP4 hardware. Technique: train on high‑fidelity simulations, prune, quantize to FP4 / NVFP4 (4‑bit) and compile for the device. Reported results: negligible accuracy loss for fluid and microclimate forecasts with empirical speedups of 10–100× (typical 15–20×), demonstrated on drone navigation and pipeline monitoring.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:36Physics models in daily apps
- 1:41Neural PDE solvers
- 2:45Edge computing for real-time physics
- 3:30Dell PowerMax GB10 overview
- 4:12Quantization for edge deployment
- 4:55Microclimate forecasting for drones
- 5:36Industrial fluid dynamics
- 6:19Deployment pipeline
- 7:06Performance gains from FP4
- 7:26Physics-informed neural models
- 8:09Hardware beyond FP4
- 8:31Power consumption
- 9:13Use cases in disaster response
- 9:57Q&A
related talks
48mThe Power of Edge AI: Revolutionizing Future of Real-Time, Autonomous Processing
01h 11mHidden Physics Models: Machine Learning of Nonlinear Partial Differential Equations
49mNew Computer Architectures : Explore Quantum Computers & SyNAPSE neuromorphic chips
25mConvergence of AI, Simulations and HPC
39mQualcomm Technologies and the power of Windows AI Q&A
26mLow-Precision Inference without Quality Loss...
15mBuilding Gen-AI experiences using LLM and LVM
11mEmpowering scientific innovation with HPC AI and Quantum –with Aseem Datar