<p>NVIDIA has released CUDA 13.3, with a handful of updates worth knowing about depending on what you’re building.</p>
<h2>CUDA Python 1.0</h2>
<p>After what feels like a long time in preview, CUDA Python hits 1.0 with stable semantic versioning. The release adds green contexts, process checkpointing, and IPC support, making it a more complete option for Python-based AI, data science, and scientific computing workloads.</p>
<h2>CUDA Tile for C++</h2>
<p>CUDA Tile is a new high-level programming model for writing portable GPU kernels in C++. It targets Hopper and other architectures, letting developers work at a higher abstraction level without giving up control over performance-critical code.</p>
<h2>CompileIQ</h2>
<p>This is probably the most interesting addition for people who spend time squeezing performance out of kernels. CompileIQ is a compiler auto-tuning framework that uses evolutionary and genetic algorithms to search for better compilation configurations. NVIDIA reports up to 15% speedups on GEMM and attention kernels. Whether that holds up across different workloads remains to be seen, but the approach is worth watching.</p>
<h2>Other updates</h2>
<ul>
<li><strong>C++23 support</strong> in NVCC and NVRTC</li>
<li><strong>CCCL 3.3</strong> adds DLPack and mdspan tensor interoperability, plus new parallel algorithms including FindIf, segmented scan, and binary search</li>
<li><strong>Random number distribution library</strong> is now more comprehensive</li>
<li><strong>Numba CUDA MLIR backend</strong> is new, with roughly 1.4x faster JIT compile times</li>
<li><strong>Math libraries</strong> (cuBLAS, cuSPARSE, cuSOLVER) get performance improvements for Blackwell and Hopper</li>
<li><strong>mmap() support</strong> added</li>
</ul>
<p>The CompileIQ and CUDA Python 1.0 milestones are the headline items here. The rest is incremental but useful, especially the Numba JIT improvements for anyone doing rapid kernel iteration in Python.</p>


Collections

NVIDIA has released CUDA 13.3 with several notable updates. CUDA Python reaches 1.0 with stable versioning, adding green contexts, process checkpointing, and IPC support for Python-based AI and scientific computing. CUDA Tile introduces a high-level C++ programming model for portable GPU kernels targeting Hopper and other architectures. CompileIQ is a new compiler auto-tuning framework using evolutionary and genetic algorithms, reporting up to 15% speedups on GEMM and attention kernels. Additional updates include C++23 support in NVCC/NVRTC, CCCL 3.3 with DLPack and mdspan tensor interoperability, a new Numba CUDA MLIR backend with ~1.4x faster JIT compile times, and performance improvements to math libraries for Blackwell and Hopper architectures.

NVIDIA CUDA 13.3: CUDA Python 1.0, Tile programming for C++, and CompileIQ autotuning