Perhaps the most significant change in CUDA’s history, since its inception, is the introduction of the programming model. Originally debuted in CUDA 13.1 and expanded in 13.2, this moves developers away from managing thousands of individual low-level threads (the SIMT model) to working with high-level "Tiles" of data.
Introduced the "largest update in two decades," featuring NVIDIA CUDA Tile , a tile-based programming model that abstracts specialized hardware like Tensor Cores. cuda driver release news exclusive
CUDA Driver Release News Exclusive: Inside the NVIDIA 13.x Ecosystem and Next-Gen Architecture Evolution Perhaps the most significant change in CUDA’s history,
[Old Architecture] GPU 1 Memory -> PCIe Bus -> CPU Controller -> PCIe Bus -> GPU 2 Memory [CUDA VMM v2] GPU 1 Memory ----------------(Direct Peer-to-Peer)----------------> GPU 2 Memory CUDA Driver Release News Exclusive: Inside the NVIDIA 13
Enhanced hardware-scheduled asynchronous copy engines allow overlapping data transfers with tensor core execution pipelines more efficiently.
Official support for the latest C++ standard brings modern features to GPU programming.