Detailed Project Report · Hardware Accelerator · ASIC · May–Aug 2026

LSTM Accelerator: RTL-to-GDSII ASIC Implementation

Qinyan Zeng
Department of Electrical and Computer Engineering
University of Southern California · qyzeng.work@gmail.com

Abstract

The accelerator implements four LSTM gates using a 16-lane fixed-point datapath, signed 8×8 MAC operations, state registers, LUT-based activation approximation, and an FSM for computation and handshaking. SystemVerilog simulation is checked against Python-generated reference vectors before physical implementation in the GF180MCU LibreLane/OpenROAD flow.

Keywords—SystemVerilog, Python, LibreLane, OpenROAD, GF180MCU, STA.

I. Project summary

Parallelism16 lanes
Number formatINT8 / Q1.7
MACSigned 8×8
Post-route frequency10 MHz
Physical resultDRC/LVS clean

II. Scope and contribution

  • Designed the datapath, state storage, and FSM controller.
  • Implemented sequential f, i, g, and o gate computation with LUT activations.
  • Generated Python golden vectors and built the SystemVerilog testbench.
  • Ran the GF180MCU RTL-to-GDSII flow.

III. Methodology

  1. Fixed-point arithmetic
  2. Sixteen compute lanes
  3. LUT activation approximation
  4. Golden-model comparison
  5. Open-source physical design flow

IV. Verification

  • Python reference-vector comparison.
  • Sequencing, arithmetic, state-update, and handshake checks.
  • Post-route timing analysis.
  • Final DRC and LVS.

V. Reported results

  • Achieved the reported 10 MHz post-route result.
  • Generated DRC/LVS-clean GDSII.

VI. Artifacts and data availability

The résumé establishes the following artifact classes. Public-safe figures and report excerpts will be attached after review:

  • Architecture
  • RTL hierarchy
  • Golden-vector flow
  • Simulation waveforms
  • Routed layout
  • Signoff summary