Skip to content

Course in preparation · Coming soon

GPU Kernels: Triton and CUDA, Measured

Write the operations a model spends its time in, and prove they are faster. Profile a model, write the hot operation as a Triton kernel and as CUDA, and check each against the roofline.

Who it is for. Engineers headed for inference and performance teams, where postings name CUDA, Triton and kernel work.

The planned syllabus

Planned lessons will help you understand each method, apply it to an AI problem, and judge whether the result holds up.

  1. 01

    The GPU memory hierarchy

    Threads, warps, shared memory and bandwidth, and the roofline that bounds any kernel.

  2. 02

    Your first Triton kernel

    Vector add, then softmax, each tested against PyTorch and timed honestly.

  3. 03

    Matrix multiply and tiling

    Why tiles matter, measured as achieved bandwidth and FLOPs on Colab's GPU.

  4. 04

    Fused attention

    Fuse the attention steps and measure the memory traffic you saved.

  5. 05

    CUDA when Triton is not enough

    The same kernel in CUDA, and when the extra control pays.

Start now with what it builds on

All courses