Announcements

To be on our mailing list, join the Google Group at this link


Aug 20, 2026: CUDA Tile C++

Time and Location

THURSDAY, Aug 20, 2026 at 6:30 PM to 8:00 PM

In Person:
Bellevue City Hall, Room 1E-121
450 110th Ave NE
Bellevue, WA 98004
Directions
Free public parking is available from 110th Ave NE, and marked “City Hall Parking”. Additional free parking is located along the light rail line.
Bellevue City Hall is located adjacent to the Bellevue Transit Center for those wanting to take public transportation!

Remote:
On Line Using Microsoft Teams
Click here to join the meeting

Title

CUDA Tile C++

Abstract

CUDA Tile C++ is a novel C++ DSL and compiler that automatically parallelizes array operations across GPU threads while presenting a single logical thread to the programmer. In CUDA Tile C++, parallelism is expressed through SIMD-style operations on arrays of data called “tiles” while memory accesses are written using structured memory abstractions. The compiler analyzes the tile and memory operations to generate an equivalent multi-threaded program that leverages the advanced hardware capabilities of modern GPUs.

In this talk, we will recap the traditional “Single Instruction Multiple Thread” programming paradigm in which the GPU programmer must manage an increasing variety of hardware features from shared memory to Tensor Memory Accelerators. We will then explore how GPU code can be made portable, simple, and performant by leveraging three key abstractions of Tile C++:

  • cuda::tiles::tile - a multi-dimensional array with value semantics whose operations are parallelized across the GPU
  • cuda::tiles::tensor_span - a view into a multi-dimensional in-memory array with an interface similar to std::mdspan
  • Tile Views - adapters that present a uniform “chunking” of a tensor_span which gives the compiler visibility into memory access patterns

These abstractions allow the Tile C++ compiler to schedule the program across hardware threads all while preserving the illusion of a single-threaded abstract machine to the user. Most importantly, code written for Tile C++ closely models the programmer’s intent and can therefore be retargeted to newer GPU architectures with substantially different hardware capabilities.

Speaker Bio

Ezra Stein is a Senior Compiler Engineer working on NVIDIA’s CUDA C++ Frontend team. Since joining the company in 2025, Ezra has focused primarily on the design and implementation of CUDA Tile C++. Prior to working at NVIDIA, Ezra spent 5 years developing MATLAB to C++ source-to-source compiler technologies at MathWorks. Ezra holds a bachelor’s degree in Computer Science and Mathematics from Cornell University.

Resources

Coming Soon!