Tensor Programs → GPU Kernels, Explained. Part 2
Scheduling and Code Generation · How compiler decisions shape storage, GPU kernels, and hardware limits.
Read more: Tensor Programs → GPU Kernels, Explained. Part 2
AI/ML inference, compilers, kernels, and accelerators.
2025–present
Building compiler infrastructure, graph optimizations, GPU and NPU kernels, and the runtime for efficient model execution. Bringing Modular to Snapdragon.
2019–2025
Worked on the heterogeneous inference engine for Apple Silicon, Neural Engine compilers, and Core ML compiler infrastructure and optimizations for Siri, Apple Foundation Models, and open‑weight LLMs.
2013–2015
GPU parallel computing research with John Owens. Co-authored Gunrock, a CUDA graph analytics library whose paper won Distinguished Paper at PPoPP 2016.
Earlier, I worked on distributed training (e.g., “ImageNet in 1 hour”) at NovuMind and on cloud services at IBM Silicon Valley Lab.
Scheduling and Code Generation · How compiler decisions shape storage, GPU kernels, and hardware limits.
Read more: Tensor Programs → GPU Kernels, Explained. Part 2Semantic Graphs · One linear layer, four ways to express the computation.
Read more: Tensor Programs → GPU Kernels, Explained. Part 1A shared vocabulary for on-device NPUs: hardware components and low-level programming concepts across five NPU families, with sources and a downloadable poster.
Read more: On-Device NPU Terminology Cheat SheetIf CUDA Graphs already cut kernel launch overhead, why build a megakernel? One measured H100 workload compares graph replay, local fusion, and keeping tiles on-chip.
Read more: CUDA Graphs, Fusion, and MegakernelsA shared vocabulary for data-center accelerators: NVIDIA and AMD GPUs, Google TPU, and AWS Trainium, compared term by term with primary sources and a downloadable poster.
Read more: AI Accelerator Terminology Cheat Sheet