Triton: Open-Source GPU Kernel Compiler | Generated by AI

Home 2026.06

Here’s what Triton is:

Triton (v3.8.0 in your checkout) is an open-source language + compiler for writing high-performance GPU kernels. Created by OpenAI (now under triton-lang org), it sits between CUDA and Python — you write kernels in Python-like syntax, and the compiler generates optimized GPU machine code (PTX for NVIDIA, AMDGCN for AMD).

Core idea: Instead of manually managing threads, shared memory, and synchronization like in CUDA C++, Triton lets you express computations at the tile level. You declare pointer blocks, and the compiler handles tiling, memory coalescing, and register allocation for you.

Architecture (what’s in the repo):

  1. Frontend (Python) — python/triton/language/core.py defines the DSL: tl.load, tl.store, tl.dot, tl.program_id, tensor operations. The @triton.jit decorator JIT-compiles Python functions into GPU kernels.

  2. Compiler (MLIR-based, C++) — The heavy lifting:
    • lib/Dialect/Triton/ — Triton IR dialect (the high-level tile ops)
    • lib/Dialect/TritonGPU/ — TritonGPU IR (GPU-specific layout/scheduling)
    • lib/Conversion/TritonToTritonGPU/ — Lower tiles → GPU-specific ops
    • lib/Conversion/TritonGPUToLLVM/ — Lower to LLVM IR → PTX/AMDGCN

  3. GPU backends — third_party/nvidia/ (NVIDIA SM80+, WGMMA, tensor memory) and third_party/amd/ (ROCm, MFMA, WMMA)

  4. Runtime — python/triton/runtime/ — JIT compilation, autotuning, caching, kernel launch

  5. Proton — third_party/proton/ — GPU profiler/diagnostic tool

  6. Gluon — python/triton/experimental/gluon/ — experimental lower-level DSL for fine-grained memory layout control

What you’d actually write:

import triton
import triton.language as tl

@triton.jit
def add_kernel(X, Y, Z, N, BLOCK: tl.constexpr):
    pid = tl.program_id(0)
    offsets = pid * BLOCK + tl.arange(0, BLOCK)
    mask = offsets < N
    x = tl.load(X + offsets, mask=mask)
    y = tl.load(Y + offsets, mask=mask)
    tl.store(Z + offsets, x + y, mask=mask)

This compiles down to optimized PTX without you touching shared memory, warp scheduling, or memory coalescing logic.

Why it matters for your trajectory: Triton is the bridge between Python-level ML code and GPU hardware. Understanding the compiler pipeline (TTIR → TTGIR → LLVM IR → PTX) is exactly the kind of deep AI infra knowledge that separates ML engineers from AI engineers. It’s also where Karpathy-style first-principles thinking meets real hardware — you need to understand memory hierarchies, warp-level primitives, and GPU architectures to write good Triton kernels.

Back Donate