| | Recommended Action | |---|---| | All users (critical) | Update to Windows driver 569.49+ or Linux driver 590.48.01+ immediately to patch CVE-2026-24187 | | Data center operators | Validate and deploy R580 LTS branch (580.126.20 or newer) for CUDA 13 workloads, with three years of support through 2028 | | Legacy GPU users (Maxwell/Pascal/Volta) | Stay on CUDA Toolkit 12.9 and Driver branch 580; CUDA 13+ drops offline compilation support for pre-7.5 compute capability | | Hopper H100/H800 users using tensor core sparsity | Monitor upcoming R535/R580 updates for the silent data corruption fix | | AI/ML developers | Adopt CUDA 13.2 with CUDA Tile for Blackwell and Ampere GPUs; leverage NIXL in CUDA DL containers for cross-node optimization | | Performance-sensitive deployments | Upgrade to CUDA 13.0 Update 1 (minimum) for FP4 GEMM, SYMV, and kernel launch latency improvements |
Buried in the R570 driver package is a new header file: cudaDriverExtension.h . It exposes three new functions that have never been publicly documented:
NVIDIA credited several external security researchers for responsibly reporting the flaws, including researchers from Seoul National University and Binarly Research Team. cuda driver release news exclusive
First‑class tile programming for C++ developers, with support expanded to Compute Capability 9.0 (NVIDIA Hopper) GPUs and all other supported architectures.
Prior CUDA updates focused primarily on optimizing specific library functions or introducing minor compiler flags. This release re-engineers the runtime environment to maximize the throughput of next-generation tensor cores. Key advancements include: | | Recommended Action | |---|---| | All
NVIDIA Explodes AI Performance with Next-Gen CUDA Driver Architecture
The most recent update for the CUDA platform is the release of CUDA Toolkit 13.2 Update 1 , which became available on April 12, 2026 . This update is a critical follow-up to the major Prior CUDA updates focused primarily on optimizing specific
18;write_to_target_document1b;_p7DsabywN4CcptQPrKK9oQg_100;57; 0;98f;0;61d;
At GTC 2026 (March 16, 2026), Jensen Huang marked the , describing it as the "flywheel" driving accelerated computing and supporting "every single phase of the AI lifecycle". He detailed the massive scale: billions of GPUs running CUDA globally form the base that attracts developers creating new algorithms.
CUDA is evolving to treat the entire data center as a single computer, requiring three core capabilities: (consistent identifiers across all nodes and GPUs), multi-node CUDA Graph (single-point launch across the entire data center with strong dependency constraints), and global memory management (cross-node unified memory views with fine-grained visibility control).