TIRx @GPUMODE

TIRx @GPUMODE

Uploaded July 2026 | Updated September 2026, 2 hours ago

Bohan Hou introduces TIRx, an open-source programming DSL for GPUs and AI accelerators that combines native-level control with tile APIs that reduce boilerplate, built on a simple layout system.

TIRx is being used to build frontier ML kernels and mega-kernel systems while exploring new approaches to agentic kernel programming.

TIRx

Lecture 52: Scaling Laws for Low Precision

Lecture 66: Game Arena

Proving Kernels Correct Instead of Testing Them

How FlashAttention 4 Works

Lecture 20: Scan Algorithm

Livestream: Outperforming cuBLAS on H100

Mirage (MPK): Compiling LLMs into Mega Kernels

Lecture 114: PyCuTe

Learning CUTLASS the hard way

Lecture 63: Search-Based Deep Learning Compilers

Lecture 89: cuTile (from friends at NVIDIA)

TIRx