pulsario.Get started
THE CPU / GPU TRACE EXPLORER LOCAL BY DESIGN

Every cycle
tells a story.

Explore CPU scheduling, GPU execution, and the performance
bottlenecks between them—all in one timeline.

Get Started Explore Features
pip install pulsario
PYPI RELEASE UPCOMING

Explore a recorded session

Pulsario
TimelineCPUSchedulerGPUFlamegraphSession
Unverified alignmentClock relation not established. Cross-source causal analysis is disabled.
Inspector
Overview
Recorded time
0.0000 s0.0050 s0.0100 s0.0150 s0.0200 s
⌄CPU · example.perf.data
main [1200]36
worker [1201]24
⌄GPU 0 · Context 1
GPU activity4
All kernels3
Memory operations2
⌄Host · main [1200]
CUDA synchronization (host waits)1
CUDA API6
NVTX2
CPU ticks are point samples. Overlapping kernels use concurrency lanes, not stream lanes.
CPU + GPUperf 20.000 msnsys 20.000 msUnverified alignmentview 20.000 ms
Selected vector_reduce_f32, duration 3.840 ms
Interactive mockup · entirely synthetic data · recorded, not liveClick events to inspect. Filtering, zoom, and GPU grouping work; other app chrome is illustrative.
FROM THE SYSTEMS
YOU WORK WITH.
Linux perf NVIDIA Nsight Systems CUDA
01 / SEE THE WHOLE PICTURE

Less context switching.
More context.

Your CPU and GPU don’t work in isolation.
Your analysis shouldn’t either.

ONE CONNECTED VIEW

Two processors. One timeline.

Bring Linux perf and Nsight Systems reports into the same workspace. Follow host-side CUDA calls to the kernels they launch, with clock alignment made explicit.

Explore combined analysis
FOLLOW THE GPU

Every kernel, in context.

Inspect CUDA APIs, kernel execution, memory copies, and GPU streams. Find launch latency and the idle gaps between useful work.

Explore GPU analysis
⌘
GO BELOW THE SURFACE

From hot paths to call stacks.

Open perf.data to explore sampled CPU threads. Follow symbolized call stacks and flamegraphs to the functions that deserve a closer look.

Explore CPU analysis
⌁
YOUR WORKSTATION. YOUR DATA.

Local is a feature.

Analyze existing traces offline in your browser. No account, no upload service, no trace data leaving your machine. Collect on ARM64 Linux; inspect on macOS or Ubuntu.

See supported platforms
IN DEVELOPMENT

A closer look at the scheduler.

Context-switch records are decoded today. Full scheduler analysis, running / runnable / blocked tracks, and runqueue latency are planned; scheduler tracepoint payloads are not yet decoded.

02 / START WITH YOUR TRACE

Bring the files.
Find the story.

Keep the profiling tools you know.
Pulsario analyzes existing files.
It never runs profiling collectors.

01

CPU only

perf.data↓Pulsario

Explore CPU samples, hot functions, and call stacks.

02

GPU only

.nsys-rep↓Pulsario

Follow CUDA calls, GPU kernels, and stream activity.

03

Better together

perf.data + .nsys-rep↓Unified timeline

Connect host and device work with explicit clock alignment.

Opening .nsys-rep requires a local Nsight Systems installation. Already-exported .sqlite files open directly.

03 / FROM TARGET TO WORKSTATION

Capture there.
Understand here.

Analyze NVIDIA Nsight Systems profiles from GPU workstations, Jetson, DGX Spark, and DRIVE AGX. Collect with your profiling tools, then bring the recorded files to your macOS or Ubuntu workstation.

Platform requirements
Your profiling targetNVIDIA GPU · LINUX CPU
perf.data / .nsys-rep
····································→
Your workstationmacOS / Ubuntu
THE DETAILS MAKE THE DIFFERENCE.

Find your next
“there it is.”

Get to know Pulsario while the public release is on its way.

Read the documentation Free to use. Proprietary. Built for performance engineers.