// tag
#cuda
2 posts tagged #cuda.
CUDA lock-in is real, and it is not in the C++
The usual account says CUDA's warp primitives and template layers weld your code to Nvidia hardware. Compile the same kernel through a third-party toolchain for an AMD GPU and the warp shuffle comes out as native cross-lane instructions with the addition fused in. What binds you is the toolchain and the licence, not the language.
CUDA 13.3: tile programming in C++ without the boilerplate
NVIDIA CUDA 13.3 (May 26) adds C++ tile programming: declarative tile abstractions replace manual shared memory, synchronization, and indexing. CompileIQ autotuning uses evolutionary algorithms to tune tile sizes and memory layout per kernel (up to 15% speedup on GEMM/attention). Works on Hopper and all other supported architectures.