Utilizing AMD Instinct GPU Accelerators for Weather and Precipitation Forecasting with NeuralGCM

2026-03-19 · Source: AMD ROCm Blogs · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, long

Summary

AMD has detailed how to utilize AMD Instinct GPU Accelerators for weather and precipitation forecasting using NeuralGCM, an innovative hybrid model. NeuralGCM integrates General Circulation Models (GCMs) with Machine Learning (ML) to overcome limitations of both traditional GCMs, which suffer from computationally expensive parameterizations and structural errors, and pure ML models, which often produce overly smooth long-term forecasts and lack uncertainty estimates. The NeuralGCM framework, designed to be fully integrated and differentiable, uses a Learned Encoder and Decoder for data conversion, a JAX-based Dynamical Core to solve primitive equations, and a Learned Physics Module (a Multi Layer Perceptron) to replace traditional parameterizations. Inference on an AMD Instinct MI300X GPU for a 4-day forecast takes 110 seconds for the 0.7° model, 48 seconds for 1.4°, and 33 seconds for 2.8°, with predictions closely matching ERA5 ground truth data.

Key takeaway

For AI Engineers developing weather or climate models, NeuralGCM offers a robust hybrid approach that mitigates the weaknesses of pure GCMs and ML models. You should consider integrating differentiable ML components into physics-based simulations to achieve more stable and accurate long-term predictions, especially when deploying on AMD Instinct GPU Accelerators for optimized performance.

Key insights

NeuralGCM combines GCMs with ML for stable, accurate, and physically consistent weather and precipitation forecasts.

Principles

• Hybrid models integrate physics-based GCMs with ML for enhanced accuracy.
• Differentiable frameworks enable ML to learn within physical laws.
• ML can replace computationally expensive GCM parameterizations.

Method

NeuralGCM employs a Learned Encoder/Decoder for data, a JAX-based Dynamical Core for fluid dynamics, and a Learned Physics Module (MLP) for sub-grid processes, trained online for stability.

In practice

• Run NeuralGCM inference on AMD Instinct GPUs using Docker.
• Utilize pre-trained NeuralGCM checkpoints for various resolutions.
• Compare forecast outputs against ERA5 ground truth data.

Topics

• NeuralGCM
• Weather Forecasting
• AMD Instinct GPUs
• Hybrid AI Models
• General Circulation Models

Code references

• ROCm/rocm-jax

Best for: Machine Learning Engineer, AI Engineer, MLOps Engineer

Related on AIssential

• Fine-Tuning AI Surrogate Models for Physics Simulations with Walrus on AMD Instinct GPU Accelerators — Fine-tuning foundation models like Walrus significantly improves physics simulation accuracy and efficiency on AMD GPUs.
• How NVIDIA's Earth-2 uses AI to Accurately Predict Weather — NVIDIA Earth-2 provides an open, GPU-accelerated AI stack for faster, more accurate weather and climate forecasting.
• From Short Histories to Long Futures: Horizon-Aware Graph Neural Networks for Long Horizon Forecasting — A multi-horizon GNN emulator improves long-range geophysical forecasting by learning state transitions and predicting increments.
• GROMACS on AMD Instinct GPUs: A Complete Build Guide — Optimizing GROMACS on AMD Instinct GPUs requires building a GPU-aware communication stack from source.
• GPU-Accelerated Deep Learning for Heatwave Prediction and Urban Heat Risk Assessment — GPU-accelerated ConvLSTM models predict next-day urban thermal conditions and heat risk by integrating satellite and meteorological data.
• The weather and climate science AI revolution isn’t revolutionary — Machine learning enhances weather and climate modeling through efficiency and targeted applications, not by replacing core physics.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AMD ROCm Blogs.