Tony Kim
Aug 05, 2025 04:45
NVIDIA plans to remove support for offline compilation of GPU architectures before compute capability 7.5 in the upcoming CUDA 13.0 release, affecting older GPU models such as Maxwell, Pascal, and Volta.
The impending release of NVIDIA’s CUDA 13.0 Toolkit will bring significant changes to GPU architecture support, particularly affecting older models. NVIDIA has announced that the new version will no longer support offline compilation for architectures predating compute capability (CC) 7.5, which includes the NVIDIA Maxwell, Pascal, and Volta architectures.
NVIDIA Software Stack
The software stack for NVIDIA’s GPUs comprises two key components: the CUDA Toolkit and the NVIDIA Driver. The CUDA Toolkit is a software development kit (SDK) that provides the necessary tools for programming GPUs, while the NVIDIA Driver acts as an interface between application software and GPU hardware. The driver is crucial for the seamless operation of applications as it directly interacts with the GPU hardware.
Driver Lifecycle and Branches
NVIDIA releases its drivers through three distinct branches: New Feature Branch (NFB), Production Branch, and Long-Term Support Branch (LTS). NFB drivers are released frequently and are suited for developers seeking the latest features. Production Branch drivers, updated biannually, are intended for enterprise and data center use. LTS drivers, updated every 18 months, offer extended support, making them ideal for long-term projects.
Implications for GPU Support
The upcoming removal of offline compilation support for architectures before CC 7.5 is part of NVIDIA’s routine support process, which phases out older architectures as newer ones are introduced. This change will affect GPUs with CC 5.x, 6.x, 7.0, and 7.2, rendering them unsupported for new application development using NVCC from CUDA 13.0 onwards. However, applications built with earlier versions of the CUDA Toolkit will remain operational on existing hardware supported by the driver.
The NVIDIA driver branch 580, set to release alongside CUDA 13.0, will be the last to support these older GPUs. As an LTS driver, it will be supported until 2028, ensuring continued operation of applications targeting these architectures.
Guidance for Developers
Developers who wish to continue supporting GPUs before CC 7.5 should maintain their development environments on the NVIDIA driver branch 580 and CUDA Toolkit 12.9. This combination will preserve the ability to compile applications for older architectures while receiving necessary updates and support through 2028.
For more detailed information on these changes, visit the NVIDIA Developer Blog.
Image source: Shutterstock
#NVIDIA #Announces #Future #CUDA #Toolkit #Support #Impacting #Older #GPU #Architectures