MLBuf-RePlAce: open-source learning-based virtual buffering for global placement
Researchers present MLBuf-RePlAce, an open-source framework that uses recursive learning to predict buffers during global placement, reporting large timing improvements in OpenROAD and commercial flows.
In plain words
Chip designers must decide where to place small amplifier cells called buffers so signals arrive on time, but doing this early in layout is computationally hard. MLBuf-RePlAce uses a machine learning model to predict buffer types and locations during the global placement step. The authors report better timing results than the default OpenROAD approach without increasing routed power.
What was announced
The arXiv paper, by the MLBuf-RePlAce authors, argues that because wire delay scales worse than cell delay in modern technology nodes, placement must account for buffer porosity, meaning cell density, to close timing in physical synthesis. It identifies two gaps in prior work: classical van Ginneken-Lillis-style buffering is too expensive to run during global placement, and learning-based methods such as BufFormer do not fully handle Electrical Rule Check (ERC) violations or feed results back into the physical design flow.
MLBuf-RePlAce is described as the first open-source, learning-driven virtual buffering-aware analytical global placement framework, built on OpenROAD. It uses recursive learning-based generative buffering to predict buffer types and locations while addressing ERC violations during placement. Against OpenROAD's default virtual buffering-based timing-driven global placer, using open-source testcases from the TILOS MacroPlacement and OpenROAD-flow-scripts repositories, the authors report maximum and average total negative slack improvements of 56% and 31% in the open-source flow without post-route power degradation, and 53% and 28% in a commercial flow with 0.2% average power improvement.
What it does not show
- The evidence is the paper's abstract only; full methodology, experimental setup and peer review are not covered, and the reported timing and power figures are author claims not independently confirmed.
For security and IT teams
- Scope
- Results cover global placement buffering evaluated on open-source testcases within OpenROAD and a commercial flow, as described in the abstract; broader production applicability is not established in this evidence.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.