DGX Spark Live: cuTile Kernels from Spark to Cloud

CUDA Tile (cuTile) is a new way to program for the CUDA platform — the bedrock upon which modern AI has been built. cuTile helps developers make use of powerful GPU features like Tensor Cores without needing to customize their code to the specifications of a particular GPU. With cuTile Python, developers can write these GPU kernels natively in Python, focusing on their algorithms while the compiler handles the hardware mapping. In this demo, we showcase cuTile in action by replacing three performance-critical kernels in the Qwen 2 7B model with custom cuTile Python implementations. The modified model is first developed and validated on a DGX Spark, then deployed without any code changes to a cloud-based B200. The CuTile compiler automatically adapts to the capabilities of each target GPU, and performance scales accordingly.