Skip to content

Welcome to Stream

Stream is a design space exploration (DSE) framework for mapping deep neural networks (DNNs) onto multi-core heterogeneous dataflow accelerators. It supports a wide spectrum of architectural designs and scheduling granularities—from traditional layer-by-layer execution to advanced layer-fused processing—enabling scalable, efficient deployment of modern DNNs.

Stream builds upon the ZigZag framework but significantly extends its capabilities to multi-core and fine-grained scheduling contexts.

✨ Key Features

  • Layer Fusion Support
    Enables splitting and scheduling parts of layers across multiple cores for higher utilization and lower memory access costs.

  • Heterogeneous Multi-Core Scheduling
    Models realistic accelerator architectures including cores with different compute/memory capabilities and interconnects.

  • Memory & Communication-Aware Analysis (COALA)
    Stream integrates COALA: a validated latency and energy model that captures data reuse, communication overhead, and memory hierarchies.

  • Workload Allocation via Constraint Optimization (WACO)
    Stream includes a built-in engine that explores valid allocations across cores using constraint-based optimization.

  • Validated Against Real Hardware
    Performance models and predictions are benchmarked against three state-of-the-art accelerator designs.

  • Modular & Extensible
    Stages of the mapping process are customizable, enabling easy experimentation and research integration.

🚀 Get Started

  1. Clone and install requirements
git clone https://github.com/KULeuven-MICAS/stream.git
cd stream
pip install -r requirements.txt
  1. Try the tutorial
git checkout tutorial
python lab1/main.py

More step-by-step setup help can be found in the Getting Started and Installation pages.

📚 Publications

The framework and methodology are described in:

A. Symons, L. Mei, S. Colleman, P. Houshmand, S. Karl and M. Verhelst,
“Stream: Design Space Exploration of Layer-Fused DNNs on Heterogeneous Dataflow Accelerators”,
IEEE Transactions on Computers, 2025.
📄 Read our paper

Stream enables researchers and developers to design, evaluate, and optimize novel DNN hardware accelerators — particularly for latency-sensitive, power-constrained edge applications.

Happy exploring!