AI/GPU-Accelerated Place&Routing Engine

Features

Place

Global Route

Timing Analysis

Physical Optimization

End-to-end AI/GPU empowerment

Place

  • GPU accelerated
  • Timing driven
  • Routability driven
  • Region & density constraints
  • AI-powered multi-objective optimization

Optimization

  • GPU-accelerated STA with high correlation to golden tools
  • GPU-accelerated Global Router and RC estimation
  • GPU-accelerated optimization operators
  • AI-powered design space exploration

Open & Adaptive

  • Full PDK support down to 7nm
  • Extensible python API with interactive shell and user-defined function
  • Full observability & controllability via Python API, seamlessly integrate with any AI workflow

Our Vision
Agile, AI-Augmented Design Flows for Faster, Smarter Chips

We envision a future where digital chip design is no longer fragmented across siloed tools and manual iterations. By unifying AI/GPU-accelerated physical implementation with other front-end engines, enable a truly AI-Augmented, agile design flow. Further, intelligent AI Agents that autonomously explore the design space, co-optimize synthesis and placement decisions, and continuously learn from past projects.

Our mission is not to merely accelerate existing workflows, but to fundamentally reinvent how chips are designed — making intelligence, integration, and acceleration the new standard for digital implementation.

Contact us

Loading...

After submission, we will contact you via email within 5 business days. Please ensure your inbox accepts emails from external senders to avoid delivery issues.

FAQ

Q: Install requirements

A:

  • Linux x86-64 operating system
  • NVIDIA GPU with compute capability >= 5.2
  • CUDA driver installed
  • HeteroPlace has been tested and is confirmed compatible with driver versions from 520.56.06 up to 530.30.02

Q: What input files does HeteroPlace require, and can it be integrated with mainstream APR tools?

A:

Yes, HeteroPlace is designed to integrate seamlessly with industry-standard tools, enabling agile physical implementation flows.

It accepts the following standard input formats:

  • LEF / TLEF – Technology and cell abstracts
  • DEF – Design layout and placement information
  • Netlist (Verilog) – Structural gate-level description
  • Liberty Library (.lib) – Timing and power characteristics of standard cells
  • SDC – Timing, clock, and design constraints
  • RC Configuration File – Layer stack and parasitic extraction rules

Q: Can HeteroPlace run on a CPU-only server (without a GPU)?

A:

Yes. While the algorithms are optimized for GPU acceleration, they also fully support multi-core CPU execution.

The CPU runtime is highly scalable and has been validated to deliver strong performance gains on systems with 128 cores or more, making it well-suited for large-scale, GPU-free compute environments.

Getting Involved

Bug report, feature request, questions are welcome! Please raise them at GitHub Issues.