CUDA 13.2 Introduces Enhanced CUDA Tile Support and New Python Features

CUDA 13.2 expands CUDA Tile support to Ampere and Ada GPU architectures (compute capability 8.x) in addition to Blackwell (10.x, 12.x), with a pip-installable cuTile Python DSL. Core runtime additions include new memcpy-with-attributes APIs, memory pool property queries, and a polymorphic cudaGraphNodeGetParams function. On Windows, GPUs now default to MCDM instead of TCC driver mode, bringing WSL2, container, and advanced memory management support. CCCL 3.2 ships modern idiomatic C++ runtime APIs (cuda::stream, cuda::buffer, cuda::launch) and new algorithms including Top-K selection (up to 5x faster than radix sort), fixed-size segmented reduction (up to 66x speedup), segmented scan, binary search, and FindIf. Python ecosystem improvements include CuPy support for CUDA 13.x, CUDA Stream Protocol interoperability with PyTorch/JAX, bfloat16 support, and cuda.core 0.6 with NVML bindings and stable CUDA Graphs API. Developer tooling gains NVIDIA Nsight Python for decorator-based kernel profiling, first-ever Numba-CUDA GPU debugging via CUDA-GDB, and Nsight Compute 2026.1 with report clustering and register dependency analysis. Embedded platforms gain MIG support on Jetson Thor for mixed-criticality workloads.