GPU profiling and observability
Profile CUDA and ROCm workloads, inspect kernels and timelines, and monitor every GPU from one browser dashboard.
Open-source clientNo credit card
Start with the workflow you need today. The same dashboard connects the results.
Profile
Trace timings, occupancy, SASS, PC sampling, and memory behavior without losing the context of the full run.
Explore profilingMonitor
Follow utilization, memory, power, and temperature across hosts, then connect a system change to the workload behind it.
Explore monitoringBuild and learn
Use Workbench for your own code or learn performance concepts through guided labs on real NVIDIA GPUs.
Open the playgroundFollow the cause
Start with the full run, open the kernel that stands out, and keep the evidence together instead of switching between disconnected captures.
Use GPUFlight with an existing CUDA application, including PyTorch workloads.
gpufl trace --upload -- python train.pyStart with one command. Add deeper profiling passes only when you need them.
GPUFlight uploads the run and organizes kernels, timelines, memory activity, and metrics in your browser.
Find the bottleneck, share the session, or compare the result with another run or host.
Need SDK instrumentation or deployment options? Read the docs.
Start with a question, work through the numbers, then test the idea on your own kernel.
Scheduling and residency
Does a larger block make my kernel faster?
Understand active warps, residency limits, achieved versus theoretical occupancy, and why a higher percentage is not always faster.
Calculate active warps and occupancyGlobal memory
Why does strided access move more data?
See how a warp turns address patterns into memory transactions and why strided access can waste bandwidth.
Compare memory access patternsShared memory
Why can shared-memory accesses serialize?
Learn why conflicting bank addresses serialize a warp request, how broadcasts differ, and when padding helps.
Map array indices to memory banksStart free, add deeper practice tools, or scale profiling to more GPUs and people.
Free
Explore GPUFlight and run occasional labs.
Practice
Build hands-on CUDA performance skills.
Individual
Profile and optimize your own workloads.
Team
Share GPU performance work with a small team.
Enterprise
Scale GPUFlight to your environment.
Credits are used for learning and profiling runs in Performance Lab and Workbench. Any credit cost is shown before you start a run. Availability and plan limits apply. CPU-only lessons and profiling on your own hardware do not use credits.
Need a custom plan? Talk to us.
Start with a free profile or explore the live demo.