← All resources

Focus on Your Algorithm—NVIDIA CUDA Tile Handles the Hardware

Auto-discovered

Blog Intermediate

Jonathan Bentz, Tony Scudiero

Visit resource ↗

Description

This is an NVIDIA developer blog by Jonathan Bentz and Tony Scudiero introducing CUDA Tile in CUDA 13.1 as a virtual instruction set for tile-based GPU programming that abstracts tensor-core details above classic SIMT.

It covers why tiles fit tensor workloads, how CUDA Tile IR complements PTX, coexistence of SIMT and tile kernels, and how most developers use cuTile Python while compiler authors target Tile IR. Figures contrast compiler-mapped tiles with programmer-mapped threads.

It is for intermediate to advanced CUDA developers and DSL authors working with tensor cores. SIMT basics and tensor-hardware motivation are assumed.

The free December 4, 2025 post includes diagrams, an explainer video, and links to docs, GitHub, and samples shipping with CUDA 13.1.

It is useful orientation before diving into cuTile samples or Tile IR specs for tensor-heavy kernels without hand-managing every tensor-core detail.