Brendan Greg (Brendan Gregg), a developer of the dynamic debugging system DTRACE, is now working on Intel and developing tools for analyzing productivity based on EBPF in the Linux kernel. Known as iaprof (ai flame Graphs), this toolkit is designed to analyze performance information of Intel GPUs and provide visualization. The code is open source and available on GitHub under the Apache 2.0 license.
The toolkit supports various hardware platforms, including Intel ARC graphic cards based on the Battlemage microarchitecture (series “B”), GPUs from the “Max” series, and Intel XE2 GPUs like IgPU Lunar Lake. It requires a Linux kernel with up-to-date Intel GPU drivers (such as Intel Battlemage needing the 6.15 core and XE driver, and Intel Max Series requiring kernel 5.15 and the I915 driver). The Linux kernel should be compiled with Intel drivers like Eu Stall and EU Debug.
The collected profiles provide insights into the performance of Execution Units, CPU states, and GPU cores. This information allows users to correlate GPU performance metrics with CPU-executed code. The toolkit is particularly useful for analyzing performance in computer games and AI applications where GPUs are intensively used, comparing GPU load with OS kernel components, runtime libraries, and AI frameworks.
Profile results can be saved in SVG format for interactive analysis in the browser with tools like flamegraph and flamescope. These interactive graphs help identify performance bottlenecks by color-coding problem areas, allowing users to drill down into stack traces, GPU instructions, and conduct detailed searches.
