Edit a CUDA project
Work with CUDA C++ source files in the browser and save each meaningful revision.
CUDA playground with built-in profiling
Run and profile CUDA code online
Write CUDA C++ or Python with Numba, compile and execute it on a managed NVIDIA GPU, and inspect the kernels, memory transfers, and timeline produced by the same run. You only need a browser—no local GPU or CUDA Toolkit.
- Real NVIDIA GPU
- CUDA C++ and Python
- Free plan, no credit card
See the workflow
From source code to a GPU profile
This short capture shows the complete loop: edit a CUDA program, run it on the managed GPU, verify the output, and open the profile created by that execution.
- EditStart with CUDA C++ or Python source.
- RunCompile and execute inside a restricted GPU sandbox.
- InspectConnect output to launches, transfers, and timing.
One connected workflow
Go beyond a basic CUDA playground
A basic CUDA playground tells you whether a program compiled and what it printed. Workbench preserves that familiar edit-and-run loop, then connects the result to GPU execution evidence. You can see which kernels launched, when memory transfers occurred, how long the work took, and what changed between source revisions.
This makes Workbench useful for focused CUDA programming exercises, small reproductions, and performance investigations. It is not a general-purpose GPU rental environment: runs execute inside a restricted sandbox and stop after a short execution window.
If you are learning performance concepts rather than bringing your own program, start with the CUDA performance guides or a guidedPerformance Lab experiment.
How it works
From source code to a GPU timeline
Keep the program, run output, and profiling evidence in one project history.
Compile and run
Execute the saved revision in a managed NVIDIA GPU sandbox and capture its output.
Inspect the profile
Connect the output to kernel launches, memory copies, NVTX ranges, and timeline data.
What you can inspect
Understand what happened on the GPU
The Trace pass captures the execution sequence behind a successful run. Paid plans can add deeper isolated passes for measured metrics, SASS, and PC sampling when the project and GPU support them.
- Kernel names, launch order, grid dimensions, and block dimensions
- Host-to-device and device-to-host memory transfers
- Per-stream GPU activity on a zoomable timeline
- Run output, errors, source revision, and profiling configuration
- Source, SASS, and sampling evidence in supported profiling modes
Choose the right starting point
Workbench or Performance Lab?
Use Workbench when you have a CUDA experiment of your own and want to connect its output to profiling data. Use Performance Lab when you want a guided problem with a target, explanation, and known optimization path. Both use the same core idea: make a small change, run it on real hardware, and verify the effect.
Common questions
CUDA playground FAQ
What languages does the CUDA playground support?
Workbench supports CUDA C++ projects and Python experiments with Numba. Available templates and profiling passes depend on the current plan and sandbox configuration.
Do I need an NVIDIA GPU or CUDA Toolkit installed?
No. Compilation and execution happen in GPUFlight's managed environment. You only need a supported browser.
Is Workbench a general-purpose cloud GPU?
No. It is designed for short CUDA experiments and profiling runs. Sandboxing and execution time limits protect the shared environment.
What does the free version include?
The free plan includes a limited Workbench project with Trace profiling. Current limits are shown on the pricing section of the homepage.
Can I inspect more than program output?
Yes. A completed profile can include kernel launches, transfers, NVTX ranges, and a visual timeline. Available deeper passes depend on the plan and workload.
Run a focused experiment
See what your CUDA code is doing on the GPU
Create a free project, run it on managed NVIDIA hardware, and inspect the result.
Open the CUDA playground