GPU Performance Engineer roadmap › Courses

Best courses for GPU Performance Engineers

This list is for people who make code run faster on graphics processors. It starts with CUDA, where you learn to write kernels, and then moves to profiling and the tools that tell you why a kernel is slow.

The documentation here is free and official. The practical habit is to measure first, change one thing, and record the result, so the speedups you claim are real.

Hiring for this role depends on code you wrote and measured. Courses set the vocabulary, but the profiler output and the benchmark log are the evidence.

  • 3 items in 2 stages: 3 courses and tutorials, 0 certifications.
  • We opened every page on November 18, 2026 and show only what the provider states.
  • No prices, ratings or enrollment counts.

See the full GPU Performance Engineer roadmap →Best books for GPU Performance Engineers →

How to choose your courses

A short method, so you spend your hours on the right thing for where you are instead of collecting certificates.

Find your stage

  • Getting started · 1 items

    CUDA fundamentals: the official programming guide and the mental model of threads, memory and parallel execution.

    Jump to this stage ↓
  • Doing the job · 2 items

    Profiling and deep learning kernels: Triton for higher-level kernels and the Nsight Compute profiler for finding the bottleneck.

    Jump to this stage ↓

If you are unsure, start one stage lower than you think. Later items assume what the earlier ones teach.

Begin with one course

Our starting pick

CUDA Programming Guide

NVIDIA

It is the official reference for the programming model you will use every day. Its early parts are written for people starting GPU programming, so you can work through them in order.

See the course ↓

Then pick by your situation

Choose the next course by the problem in front of you, not by its place in the list.

Free path or paid path?

3 of the 3 courses and tutorials can be read or taken without paying, so you can build most of the skill at no cost. Check each certification page for exam fees.

Paid courses earn their place when you need structure, deadlines or feedback. Check what the provider says is included before you commit.

Judge any course before you commit

  • Who runs it

    Prefer courses from the people who build or govern the tool. They tend to stay current, and an exam from a tool’s own maker is the clearest credential for that tool.

  • Hands-on or watch-only

    In this field you learn by running things. Choose material where you build a pipeline, a cluster or a dashboard over material you only watch.

  • How current it is

    Tools change fast. Look for a version or last-updated date on the provider’s page, and check anything several years old against the current docs.

  • What you will be able to show

    Ask what you will have at the end: a working repository, a deployed service or a passed exam. If it is only a certificate of completion, pair it with a project of your own.

Getting started

CUDA fundamentals: the official programming guide and the mental model of threads, memory and parallel execution.

  • Start hereHands-on tutorialFree to learn

    CUDA Programming Guide

    NVIDIA

    View course →

    What you will learn

    • The CUDA programming model and GPU execution
    • Programming GPUs in C++ and Python
    • Advanced CUDA features, including multiple GPUs

    Good for

    The official reference you read alongside your first kernels. Its early parts are written for anyone starting GPU programming.

    Also stated by the provider

    • Official NVIDIA documentation, updated with each CUDA release
    • Includes a PDF version

Doing the job

Profiling and deep learning kernels: Triton for higher-level kernels and the Nsight Compute profiler for finding the bottleneck.

  • Hands-on tutorialFree to learn

    Triton Tutorials

    OpenAI Triton

    View course →

    What you will learn

    • Vector addition and fused softmax
    • Matrix multiplication and fused attention
    • Layer normalization, dropout and grouped GEMM

    Good for

    Writing deep learning kernels at a higher level than CUDA, with the same memory and parallelism ideas. It is the bridge to the deep-learning kernel step.

    Also stated by the provider

    • Sphinx-generated gallery of runnable examples
    • Recommended to read in order from the simplest tutorial
  • Hands-on tutorialFree to learn

    Nsight Compute User Guide

    NVIDIA

    View course →

    What you will learn

    • Kernel profiling views and controls
    • Reading the metrics that explain a slow kernel
    • Command-line profiling for repeatable measurements

    Good for

    The profiler you will use every day in this role. Read it when a kernel is slow and you need to know which metric to trust.

    Also stated by the provider

    • Official NVIDIA user manual
    • Documentation for the current Nsight Compute release

Details such as hours, format and whether material is free are as the providers state them on November 18, 2026, and can change. Confirm on the provider’s page before you commit. See our disclosure.

Frequently asked questions

Which course should a GPU Performance Engineer take first?

CUDA Programming Guide, from NVIDIA. It is the official reference for the programming model you will use every day. Its early parts are written for people starting GPU programming, so you can work through them in order.

Can a GPU Performance Engineer learn this for free?

Largely, yes. 3 of the 3 courses and tutorials here can be read or taken without paying (CUDA Programming Guide, Triton Tutorials, Nsight Compute User Guide). Check each certification page for exam fees: not every page states one.

How was this list put together?

We opened each course, tutorial and certification page on November 18, 2026 and wrote down only what the page states: who runs it, what it covers, hours and exam format where given, and whether the material is free to read. We do not show prices, ratings or enrollment figures. Details can change, so confirm on the provider’s page before you commit.