Oriel — AI & technology worth noticing, every day.

NVIDIA CUDA 13.5 Brings NVIDIA Vera Rubin GA Support

Web page · https://developer.nvidia.com/blog/nvidia-cuda-13-5-brings-nvidia-vera-rubin-ga-support

As artificial intelligence and accelerated computing continue to push the boundaries of what's possible, the latest NVIDIA CUDA release arrives to unlock...

First seen by Oriel on .

AI summary

CUDA 13.5 makes NVIDIA Vera Rubin support generally available and expands GPU scheduling, multi-GPU CUDA Tile, Python and CCCL APIs, graph operations, and profiling and debugging tools.

  • Locality domains can favor equal SM partitions or include supplemental SMs, while MPS v3 adds weighted timeslicing, resource borrowing, cgroup routing, and stronger memory accounting.
  • CUDA Graphs now capture batched 1D copies, while cuda.core 1.2 adds copy_batch(), PDL, simpler runtime compilation, and executable-node updates without reinstantiation.
  • CUDA Tile kernels gain NVLink and NVSwitch multi-GPU communication and preview SIMT library calls, though CUDA 13.5 cannot call user-defined __device__ functions from tile code.
  • CCCL 3.5 adds public CUB tuning, batched Top-K, reproducible floating-point scans, and reductions that read element counts directly from GPU memory.
  • Rubin supports WDDM and MCDM on Windows, not TCC; Nsight Perf SDK for AI previews always-on profiling, and CUDA-GDB can stop on Driver log events.

#cuda #cuda tile #gpu scheduling #gpu profiling

Generated by gpt-5.6-sol. Check the original for details.

Aliases

SourceExternal IDURL
NVIDIA Developer Bloghttps://developer.nvidia.com/blog/?p=124464https://developer.nvidia.com/blog/nvidia-cuda-13-5-brings-nvidia-vera-rubin-ga-support/ Read discussion