The CARM Tool
The CARM Tool is a micro-benchmarking toolkit that constructs Cache-Aware Roofline Model (CARM) performance models across multiple CPU architectures (x86, ARM, RISC-V) and GPU platforms (ROCm, CUDA). It measures arithmetic performance and memory bandwidth of all cache levels to model the architecture’s performance and guide optimization.
The toolkit provides:
- Benchmarking: generates roofline plots by measuring peak performance (FLOP/s) and peak memory bandwidth for each cache level
- Profiling: profiles instrumented applications to compute roofline metrics (arithmetic intensity, FLOP/s)
- GUI: interactive Dash + Plotly dashboard for exploring roofline plots and application data
Quick Start
pip install carm-roofline
# Run a single-thread roofline benchmark (auto-detects ISA)
carm benchmark
# Launch the GUI to visualize results
carm gui
See the Installation page for detailed setup instructions.
Table of Contents
| Page | Description |
|---|---|
| Installation | Requirements, pip install, virtual environment, from-source install |
| Commands | Subcommand reference — benchmarking, GUI, profiling |