|
Listen to this story
|
NVIDIA announced the release of CUDA’s software toolkit update 12.0 that focuses on new programming models and accelerating processing capabilities through new hardware updates.
The update allows users to target architecture-specific features in the NVIDIA Hopper and NVIDIA Ada Lovelace architectures with enhanced libraries, developer tools, and CUDA custom code. The new update also supports revamped CUDA dynamic parallelism APIs, enabling performance improvements and also enhancing CUDA Graphs API.
Moreover, the update also includes support for GCC 12 host compiler and C++20. The toolkit also has a new nvJitLink library for JIT LTO. With the new update, many tensor operations are now also available through public PTX like TMA bulk operations and 32x Ultra xMMA.
Click here to check out the full release notes and here for the download links.
CUDA (Compute Unified Device Architecture) is NVIDIA’s parallel computing platform developed for general computing and is the lead framework for GPGPU. It is a software layer providing direct access to GPU’s virtual instruction set along with parallel computational elements for executing compute kernels.
In July, after the success of CUDA, NVIDIA announced the launch of their unified quantum computing platform, ‘QODA’ (Quantum optimised device architecture), for accelerating breakthroughs in quantum research and development in AI, HPC, and others.
