RTX 5090 CuPy Setup: Blackwell Architecture & CUDA 13.1

Users experiencing issues with CuPy on RTX 5090, 5080, or 5070 GPUs should note that the new Blackwell architecture requires CUDA 13.1 for compatibility. Pre-built CuPy wheels do not support the compute capability of these GPUs, necessitating a build from source. After uninstalling existing CuPy versions, install the CUDA Toolkit 13.1 and then CuPy without binaries. For Windows users, ensure the correct path is added to the system PATH. Proper configuration can lead to significant performance improvements, such as a 21× speedup in physics simulations compared to CPU processing. This matters because it highlights the importance of proper software setup to fully utilize the capabilities of new hardware.

The release of the RTX 50-series GPUs, including the RTX 5090, introduces the Blackwell architecture, which requires CUDA 13.1 for optimal performance. This is significant for developers and researchers who rely on GPU acceleration for computational tasks. The new architecture offers substantial performance improvements, but it also necessitates updates to software dependencies like CuPy, a library used for GPU-accelerated computing in Python. Without these updates, users may experience failures or suboptimal performance when running applications on these new GPUs.

CuPy, a popular library for numerical computations, does not yet provide pre-built wheels that support the compute capability 10.0 required by the Blackwell architecture. This means users must build CuPy from source to ensure compatibility with the latest NVIDIA GPUs. The process involves uninstalling any existing versions of CuPy and installing the CUDA Toolkit 13.1 before compiling CuPy manually. This extra step is crucial for leveraging the full power of the RTX 50-series, which can significantly accelerate tasks like large-scale simulations and machine learning model training.

For Windows users, additional configuration is needed to ensure the system recognizes the new CUDA version. This involves modifying the system PATH to include the appropriate directories for the CUDA 13.1 binaries. Such steps are essential to avoid runtime errors and ensure that the GPU resources are utilized efficiently. Proper setup can result in dramatic performance gains, as evidenced by a verified 21× speedup in a 1M-particle physics simulation compared to CPU execution.

The requirement to build CuPy from source highlights the importance of staying updated with the latest software and hardware developments in the field of GPU computing. As new architectures are released, they often bring both opportunities for enhanced performance and challenges in terms of compatibility and setup. Understanding these requirements and being able to adapt quickly is crucial for developers and researchers who depend on cutting-edge technology to push the boundaries of computational capabilities. This matters because it directly impacts the efficiency and feasibility of complex computational projects across various scientific and engineering domains.

Read the original article here