Skip to content

Toolchain Overview

The PLENA toolchain is a full-stack framework that takes a Hugging Face model through software optimization, simulation, and hardware evaluation. A Co-Design DSE engine drives two coupled loops: a Software Optimization Loop tuning data type and precision against accuracy, and a Hardware Optimization Loop tuning the accelerator configuration against Power, Performance, and Area (PPA).

PLENA System Architecture

Components

  • Compiler — Lowers a Hugging Face model onto the PLENA stack and emits machine code for the accelerator.
  • Accuracy Evaluator — Applies training-free quantization and measures model accuracy, closing the Software Optimization Loop.
  • Transactional Emulator — Cycle-approximate, HBM-enabled simulator that executes the generated machine code.
  • Analytic Model — Fast regression-based latency, area, and power models for rapid DSE inner-loop search.
  • Co-Design Toolchain — DSE engine that jointly explores the software and hardware design space across both optimization loops.