Close
No urgent action required unless you are affected by the open bugs. Developers targeting Blackwell B100 should remain on 12.6. Watch for CUDA 12.7 announcement at NVIDIA GTC Spring 2026 (May 18–20).
: Added support for Extended GPU Memory (EGM) arrays, improving how large datasets are handled across system and GPU memory.
: CUDA 12.x is optimized for the Hopper (H100) and Ada Lovelace (RTX 40-series) architectures, featuring better utilization of FP8 precision and improved Tensor Core performance. cuda 12.6 news today
: Introduced multi-GPU support for protected PCIe mode and key rotation for single GPU passthrough. Version History and Compatibility
Based on NVIDIA’s typical cadence, here are probable updates relevant to CUDA 12.6 today: No urgent action required unless you are affected
While gamers focus on frames per second, developers are looking at the CUDA code powering them. News from SIGGRAPH indicates that the upcoming (Deep Learning Super Sampling) and new Ray Reconstruction techniques are built heavily on the optimizations found in CUDA 12.6.
: Discussions regarding the integration of CUDA 12.6 into the NVIDIA JetPack ecosystem for Jetson devices are ongoing as of early 2026. Ecosystem Trends and Future Context : Added support for Extended GPU Memory (EGM)
: Users can now install PyTorch with CUDA 12.6 support via specific pip commands ( pip3 install torch --index-url https://download.pytorch.org/whl/cu126 ).
: New conditional execution features, including ELSE graph support for IF nodes and new SWITCH nodes, allow for more complex and efficient GPU workflows.
: Added compiler support for Blackwell SM_100, SM_101, and SM_120 architectures, including DRAM encryption query and control.
: A major shift occurred with the introduction of Range Profiling APIs in 12.6 Update 2. These APIs simplify the profiling process for new users and align with modern host-target structures, replacing legacy APIs that are now marked for potential deprecation. Math Library Enhancements :
