NVIDIA CUDA 12.2 Released With Linux HMM Support

NVIDIA CUDA 12.2 is out today and while it's just an update to the CUDA 12 series, it's actually quite an exciting release.

Most exciting with CUDA 12.2 for Linux users is this compute stack finally supporting the Linux kernel's Heterogeneous Memory Management (HMM) functionality. HMM allows for seamlessly sharing data between host memory and accelerator devices. This NVIDIA CUDA 12.2 HMM support depends on Linux 6.1.24+ or Linux 6.2.11+ for having the necessary bits. Additionally, this HMM support requires the use of the NVIDIA GPU Open Kernel Modules driver rather than the proprietary kernel driver.

This initial NVIDIA HMM support also doesn't yet work on Arm/AArch64, GPU atomic operations on file-backed memory aren't yet wired up, HugeTLBfs pages are not yet supported, and the fork() system call is not fully supported yet. NVIDIA also expects to better optimize their HMM support in future releases.

The NVIDIA CUDA 12.2 release also now enables lazy loading support by default when using the NVIDIA 535+ kernel driver, host NUMA memory allocation support, per-client priority mapping at runtime for CUDA Multi-Process Service, and a variety of fixes and other improvements.

Downloads and more details on the NVIDIA CUDA 12.2 release via developer.nvidia.com.