This release today contains more details , its stable and achieves peak performance on B200 blackwell.
Tbh the syntax is far better than python Triton for me , without all of the excessive slicing and brainfuck like incantations that need to be written between the square brackets.
With the recent changes in the GPU kernels domain , with nvidia releasing CuTeDSL in python and advising people to use it rather than cutlass c++ and urges people to start with it.
Aswell as recent updates to triton and gluon and Tilelang (all in python) how do you guys feel about rust replacing c++ for working with ML systems and interop with kernels?
Before we were mostly locked in because if you wanted gpu , you needed cuda and cutlass and aten interop with python and PyTorch, but now there are way more options and better alternatives.
Curious to hear your opinions and thoughts, personally i am rooting for rust(non bias opinion 👀) ,
Honorable Alpha mentions :
Nvidia CuTile-rs (python cutile but in rust ) and Nvidia Cuda Oxide (cuda in rust not to be mistaked with cudarc or cuda wrappers , straightcuda kernels in rust)