NVIDIA has recently released the CUDA 12.0 development tools update. The update brings a lot of changes, including support for the latest Hopper and Ada Lovelace GPU architectures, upgraded C++ dialects, official support for JIT LTO, improved and new APIs, and more.
The main features of CUDA 12.0 are described as follows
CUDA 12.0 provides programmable features for NVIDIA’s Hopper and Ada Lovelace architectures. CUDA 12 provides many tensor operation options for the Hopper and Ada architectures, many of which support the public PTX intermediate representation. Cooperative Grid Array (CGA) support for C built-in functions, programmable L2 cache for SM multi-projection, genomics/DPX instructions, and other additions.
Support for using the virtual memory management API for GPUs marked with CUDA_VISIBLE_DEVICES.
Application and library developers can programmatically update the priority of CUDA streams.
Improved CUDA dynamic parallelization API with “substantial” performance improvements over the previous API.
Instant Link Optimization (JIT LTO) is now officially supported via the nvJitLink library.
Support for the GCC 12.1 host compiler.
NVCC and NVRTC support for C++20 dialects.
NVRTC has updated its default C++ dialect from C++14 to C++17.
See the release notes for more details on the CUDA 12.0 changes. Users and developers can download CUDA 12.0 for all major platforms from developer.nvidia.com.
