AccelForge is a framework for modeling, designing, and exploring tensor algebra accelerators. It uses HWComponents as a backend for area, energy, latency, and leak power estimates.
Learn more at the website or on GitHub.
- Flexible Full-Stack Modeling of a wide variety of devices, circuits, architectures, workloads, and mappings. We integrate with HWComponents, with easily-to-modify models for component area, energy, latency, and leak power.
- Fast and optimal mapping of workloads onto architectures, yielding the best-possible performance and energy efficiency.
- Fusion-aware mapping that optimizes fusion for cascades of Einsums, enabling end-to-end optimization of entire workloads.
- Heterogenous Architectures that can include multiple types of compute units.
- Strong input validation via Pydantic, with clear error reports for invalid specifications.
- Pythonic Interfaces that enable easy automation and integration with other tools.
pip install accelforgeSee examples/ for architectures and workloads, and notebooks/ for tutorials.
If you use AccelForge in your work, please see Citing AccelForge for the relevant papers.