Posts by Baiqiang Xia
GEAK V3: Agent-Driven, Repository-Level GPU Kernel Optimization across HIP, Triton, and FlyDSL on AMD GPUs
- 20 July 2026
In the ever-evolving world of GPU computing, optimizing kernels for performance and efficiency is a critical challenge. Hand-tuning kernels demands deep technical expertise and manual iteration. In this blog, you will read about how GEAK v3, the latest iteration of the agent-driven framework, tackles this problem using enhanced features such as task planning, test-harness discovery, patch-based handling of multi-file kernels, dynamic memory system and expert knowledge database. Our results show improvements across three kernel languages (HIP, Triton, and FlyDSL) and both CDNA and RDNA GPUs.
Training AI Weather Forecasting Models on AMD Instinct
- 10 November 2025
Weather forecasting is one of the most computationally intensive scientific challenges and an essential societal need. Predicting extreme weather events, agricultural and energy planning and daily forecasts all require accurate weather predictions. Traditionally, Numerical Weather Prediction (NWP) has served as the foundation of weather forecasting by solving complex physical equations that require significant computational power. However, recent advances in machine learning have led to the development of alternative prediction models that reduce computational costs by orders of magnitude, while either maintaining or improving accuracy in forecasts. Models like GenCast [1], Pangu-Weather [2], Aurora [3] and others have shown promising results in this area (see the WeatherBench [4] scorecard). Running inference on these models using AMD GPUs is straightforward, as highlighted in our recent blog post: Running SOTA AI-based Weather Forecasting models on AMD Instinct.
Running SOTA AI-based Weather Forecasting models on AMD Instinct
- 18 September 2025
Weather Forecasting is a complex scientific problem where immense progress has been made through the Numerical Weather Prediction (NWP) approach using computational fluid dynamics-based models. Forecasting is usually done in three stages: a data assimilation stage where all available data streams at the time \(t\) (sometimes previous times can be used to improve this estimate) are used to estimate the current 3D state of the atmosphere \( S_{t}\) (surface and atmosphere), as parameterized by a number of variables at the current time \( t\), a forecasting stage where the state \(\hat{S}_{t + \delta t}\) for a later time \( t+ \delta t\) (i.e., all the variables at this later time) are forecasted, and a downstream stage where the forecasted state at time \(t + \delta t\) is used to estimate weather variables at more specific times and locations.