For those lucky enough to have H100 or B100 clusters:
For the kernel engineers writing the code that actually makes GPUs fast, the recent news regarding is the highlight. cuda release news
Faster prototyping, no context switching between Python and C++. Performance is within ~5% of hand-tuned C++ kernels. For those lucky enough to have H100 or
The biggest quality-of-life shift: are now built into the core driver API. cuda release news