GPU Algorithm Authoring with CUDA Tile
summary
Claim: tile programming with CUDA Tile and cuTile Python lets authors write array‑centric GPU kernels that stay portable across NVIDIA architectures. The session uses a browser JupyterLab on Brev, builds and runs real kernels for LLM inference and conjugate gradient, profiles and tunes with NVIDIA Nsight, and shows autotuning and tile vs SIMT tradeoffs. Measured results show tile kernels reach near state‑of‑the‑art throughput while SIMT versions run noticeably slower.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 12:37CUDA Python overview
- 23:05GPU architectures evolution
- 32:24CUDA Tile model
- 45:52Kernel implementation comparison
- 1:00:19Vector addition hands-on
- 1:13:26Performance benchmarking
- 1:23:47Q&A
related talks
28mThe Future Is Tiled: Using CuTile & TileIR To Write Portable, High-performance GPU...
31mBuilding Inference Workflows with Tile Languages
01h 00mThe Shape of Things to Come
02h 51mCUDA Python Kernel Authoring
48mGPU Programming in Pure Python
05mLightning Talk: GPU Programming with C++ and Triton
01h 29mGPU development with Python 101
30mCUDA in Your Python: Effective Parallel Programming on the GPU