DREAMPlace

The idea

Look at ePlace's inner loop with a machine-learning engineer's eyes: millions of parameters (cell coordinates), a differentiable scalar objective (smoothed wirelength plus electrostatic energy), gradient descent with momentum-like acceleration. That is neural-network training, minus the network. DREAMPlace therefore implements placement inside a deep-learning toolkit: coordinates become the trainable tensor, wirelength and density become custom differentiable operators (with hand-written CUDA (Compute Unified Device Architecture) kernels for the net-model gradients and the FFT (Fast Fourier Transform) Poisson solve), and the framework's optimizers and infrastructure do the rest.

The result is reproducible: same solution quality as RePlAce, speedups reported around thirty times and beyond on large designs, turning multi-hour global placements into minutes. And because the implementation is a differentiable pipeline, research riffs come cheap: new objectives (timing, routability terms) are new loss functions; later versions added GPU-accelerated detailed placement and legalization stages.

Why it matters beyond speed

It reframed placement culturally: EDA (Electronic Design Automation) algorithms as differentiable programs on commodity accelerator infrastructure, opening the door to hybrid schemes where learned models and classical optimization share one computation graph. Much of current machine-learning-for-placement research builds directly on this codebase.

Where you meet it today

The open-source repository is actively developed and widely forked; its ideas are visibly absorbed into vendor tools' GPU placement offerings.