What is IntelliKit?#
IntelliKit is a collection of Python tools to help you analyze and improve the performance of AMD hardware. It works with ROCm to turn data like hardware counters and traces into easy-to-use Python APIs. For host-side CPU hotspot analysis, the toolkit includes uprof_mcp, which integrates AMD uProf. Additionally, IntelliKit supports LLM-style workflows by providing MCP servers and agent skills, enabling you to use installable SKILL.md playbooks with platforms such as Cursor, Claude, Codex, and GitHub Copilot.
The following table lists all IntelliKit tools, their roles, and functionalities:
Tool |
Role |
Description |
|---|---|---|
Isolate |
Captures kernel dispatches and builds standalone reproducers for HIP and Triton code. |
|
Profile |
Translates hardware counter data into human-readable metrics like bandwidth, cache, and compute performance. |
|
Profile |
Maps GPU performance to your source code, and analyzes source-line timing and stalls. |
|
Inspect |
Intercepts HSA packets to reveal what ran on the GPU, including assembly code and HIP source details. |
|
Validate |
Confirms that optimized kernels produce the same results as their reference implementations. |
|
MCP |
Provides HIP compiler, documentation, and rocminfo servers for LLM agents. |
|
CPU |
Bridges AMD uProf for host-side CPU hotspot analysis. |
For more information about using IntelliKit tools in an end-to-end workflow, see Using IntelliKit end to end.