cuTile.jl Brings NVIDIA CUDA Tile-Based Programming to Julia

2026-03-03 · Source: NVIDIA Technical Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Emerging Technologies & Innovation · Depth: Intermediate, quick

Summary

NVIDIA CUDA Tile, a significant addition to CUDA programming, provides automatic access to tensor cores and specialized hardware. Following its Python release, cuTile.jl now brings this tile-based GPU programming model to Julia developers, simplifying high-performance kernel development. Unlike traditional CUDA, which requires explicit management of threads and memory hierarchies, CUDA Tile allows developers to describe operations on data tiles, with the compiler handling hardware mapping. cuTile.jl maintains an abstraction level identical to cuTile Python, facilitating code porting, while incorporating Julia idioms like 1-based indexing and broadcast expressions. The package targets the same NVIDIA Tile IR backend as its Python counterpart, achieving performance parity for compute-intensive kernels on NVIDIA Blackwell architecture GPUs, though some complex control flow kernels are still maturing.

Key takeaway

For Julia developers building high-performance GPU applications, cuTile.jl offers a streamlined approach to CUDA kernel development. You should consider adopting cuTile.jl to simplify complex GPU programming tasks, especially for compute-intensive kernels, as it abstracts away low-level thread and memory management. Be aware that some advanced Julia features and complex control flow kernels are still under active development and may not yet achieve full performance parity or support.

Key insights

CUDA Tile simplifies GPU programming by abstracting hardware details, enabling tile-based operations for high-performance kernels.

Principles

• Abstract hardware details for GPU programming.
• Maintain consistent abstraction across language bindings.

Method

cuTile.jl uses a custom Julia compiler to intercept standard library calls, routing them to Tile IR operations, which are then compiled to GPU machine code by the NVIDIA "tileiras" compiler.

In practice

• Use `ct.load` and `ct.store` for tile-level data operations.
• Leverage Julia's broadcasting syntax for element-wise tile operations.

Topics

• NVIDIA CUDA Tile
• Julia GPU Programming
• Tile-based Programming
• Tensor Core Acceleration
• GPU Kernel Optimization

Code references

• JuliaGPU/cuTile.jl
• JuliaGPU/CUDA.jl

Best for: Machine Learning Engineer, Deep Learning Engineer, Software Engineer

Related on AIssential

• CUDA Tile Programming Now Available for BASIC! — cuTile BASIC brings GPU acceleration to the BASIC language via CUDA Tile, simplifying parallel programming for legacy applications.
• OptiML: An End-to-End Framework for Program Synthesis and CUDA Kernel Optimization — OptiML unifies LLM-driven code generation with hardware-aware search to optimize CUDA kernels for performance.
• Evaluating CUDA Tile for AI Workloads on Hopper and Blackwell GPUs — CuTile offers high productivity and peak performance on datacenter Blackwell, but lacks portability and consistent performance across other architectures.
• NVIDIA CUDA 13.3 Enhances GPU Development with Tile Programming in C++, Compiler Autotuning, and Python Updates — CUDA 13.3 streamlines GPU development with high-level abstractions and performance optimizations across C++ and Python ecosystems.
• Nautilus: An Auto-Scheduling Tensor Compiler for Efficient Tiled GPU Kernels — Nautilus automates tensor kernel optimization from math specifications, outperforming existing compilers.
• FlyDSL: Expert GPU Kernel Development with the Ease of MLIR Python Native DSL on AMD GPUs — FlyDSL is a Python-first, MLIR-native DSL for high-performance GPU kernel development on AMD GPUs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by NVIDIA Technical Blog.