Triton Compiler Lead Update | Generated by AI

Home 2026.06

Let me search for current information on this.Question: What is Philippe Tillet (the author of Triton) doing now?

Answer:

Philippe Tillet is still at OpenAI as a Member of Technical Staff, leading the Triton compiler effort. He’s based in San Diego, with ~13 years of experience in high-performance ML systems, compiler engineering, and GPU kernel optimization. Quick recap of his trajectory and current focus:

Background. He started with GPUs in 2011 as a ViennaCL contributor, got his B.S. from Telecom SudParis (2012), M.S. from NCTU Taiwan (2014), and his Ph.D. from Harvard in 2020 with a dissertation on compilers for blocked algorithms on GPUs. He joined OpenAI full-time in 2020 to pursue Triton, a project he started in 2018 out of frustration with writing auto-tuners for matmuls in CUDA.

Current work — what’s notable for you:

  1. Triton on custom silicon. This is the most interesting recent signal. He’s expanding the Triton Compiler group at OpenAI and hiring engineers to push performance on OpenAI’s custom 1P (first-party) accelerators. That confirms OpenAI is targeting Triton as the kernel-authoring layer for their own chips (the Broadcom-partnered hardware), not just NVIDIA — a hardware-portability play that matters a lot if you’re thinking about where the MoE/inference-optimization stack is heading.

  2. Blackwell support. He presented “Blackwell Programming for the Masses with OpenAI Triton” (early 2025), covering the CUDA-C++ vs. XLA/Torch trade-off space and new TMA-style APIs like tl.make_tensor_descriptor for tiled GEMM on Blackwell. The talk gets into real hardware constraints — e.g. on H100, wgmma.sync doesn’t distinguish between QK and PV MMAs, so they can’t be pipelined independently (directly relevant to FlashAttention kernel scheduling).

  3. Research output. He has an ICML 2025 paper (ELMO) with PhD students Nasib Ullah, Jinbin Zhang, and Erik Schultheis on pure 16-bit and 8-bit training for large-output-space models, using fused Triton kernels to cut peak GPU memory from ~40GB to ~6GB on a 3M-item recommendation dataset.

Why this matters for your trajectory: The Triton-on-custom-accelerators angle is the thread worth pulling. If you’re moving toward inference optimization and GPU compute, Triton is increasingly the portable kernel IR layer — and the fact that OpenAI is betting it onto their own ASICs (the same way AMD already maintains a ROCm Triton fork, which you’d use on your MI300X) means kernel skills written in Triton compound across hardware vendors in a way CUDA doesn’t. Worth checking the Triton tutorials directly against your MI300X to see how the ROCm backend codegen differs from the NVIDIA path.

References:

Back Donate