Autodesk and NVIDIA Accelerate CFD with Warp and GH200 Superchip
Felix Pinkston Sep 16, 2025 16:52
Autodesk Research harnesses NVIDIA's GH200 Superchip and Warp framework to enhance computational fluid dynamics (CFD) performance, achieving significant speedups and scalability.
Autodesk Research has taken a significant leap in computational fluid dynamics (CFD) by utilizing NVIDIA's advanced technologies, according to NVIDIA. By leveraging the NVIDIA GH200 Grace Hopper Superchip and the Warp framework, Autodesk's Accelerated Lattice Boltzmann (XLB) library has achieved remarkable performance improvements, marking a pivotal moment for CFD simulations.
Advancements in CFD Performance
The integration of NVIDIA's Warp framework with the GH200 Superchip has enabled Autodesk to achieve an approximate 8x speedup in CFD simulations. This performance boost was measured against the GPU-accelerated JAX backend, showcasing the potential of Python-based solvers to match the efficiency of low-level programming languages traditionally used in CAE applications.
The Warp framework, open source and Python-based, bridges the gap between Python's accessibility and CUDA's high-performance capabilities. It allows developers to write GPU kernels directly in Python, which are then compiled to native CUDA code. This innovation facilitates the seamless integration of CFD solvers with modern AI and machine learning ecosystems.
Scalability and Efficiency
Autodesk Research's XLB library, powered by NVIDIA Warp, demonstrates near-linear scalability, handling up to 50 billion computational elements on an eight-node GH200 cluster. This scalability is achieved through an out-of-core computation strategy, where data is efficiently transferred between CPU and GPU memory, thanks to the GH200's NVLink-C2C interconnect.
The Warp framework's design allows for explicit memory management and optimized kernel programming, further enhancing memory efficiency and computational throughput. Compared to the JAX backend, Warp provides a two- to three-fold improvement in memory usage, making it a compelling choice for large-scale CFD simulations.
Bridging Performance and Productivity
The collaboration between Autodesk Research and NVIDIA highlights a new era where the historic tradeoff between development productivity and computational power is being overcome. The XLB library exemplifies how Python-native frameworks can deliver high-performance results comparable to optimized low-level code while maintaining the rapid development cycle associated with Python.
For more information on Autodesk Research's XLB library and NVIDIA's Warp framework, visit the official NVIDIA blog here.
Image source: Shutterstock.jpg)