Judging AI agents by the chips they lay out: Nova-SI averages 57.97 on an industrial floorplan where the strongest general configuration reaches 31.17
Synopsis
Early floorplanning decisions only reveal their consequences after downstream implementation, so successful tool execution cannot judge them. An evaluation framework connects submitted layouts to a prescribed power-grid, placement and global-routing flow, checks geometric validity, and measures congestion, timing and resources. On one industrial design whose 99 macros occupy 71.1% of the layout region, the specialized system Nova-SI reaches a mean quality score of 57.97 without online feedback, against 31.17 for the strongest general-purpose configuration, GPT-6 Astra, though their per-output score ranges overlap.
Interpretation
Chip-floorplanning agents are now judged by the physical consequences of their delivered layouts under downstream implementation, not by whether a script runs. Earlier agent evaluations targeted task completion, tool environments or inspectable deliverables, and geometric packing and wirelength proxies did not capture layout quality. On a fixed L-shaped layout with 235,727 instances and 99 hard macros covering 71.1% of the region, 39 layouts received complete physical measurements; Nova-SI averaged 57.97 in quality against 31.17 for the strongest general configuration, GPT-6 Astra.
The quality gap sits mainly in congestion and timing rather than global packing density. Mean standard-cell utilization is close between Nova-SI and Astra, at 32.27% and 33.56%. In the same comparison, mean total overflow is 17,563 versus 281,748, the affected global-routing-cell fraction is 0.0745% versus 0.9586%, and mean wirelength is 21.89 m versus 28.88 m.
The evaluation separates submission, measurement and interpretation: it first confirms that submitted macro geometry and fixed ports enter the measured state, then keeps invalid geometry apart from poor quality and missing measurements. Downstream standard-cell placement, resizing and buffer insertion legitimately change the design, so unchanged standard-cell geometry cannot be required. For 11 baselines whose submitted and measured standard-cell states differ, the audit covers all 99 macro geometries, 107 fixed signal ports and the original cell names and references, with 25 core power-grid, placement and constraint files byte-identical.
Perspective
The evaluated scope is one industrial design on a fixed L-shaped layout, with the flow stopping at power grid, placement and global-route estimation, before clock-tree synthesis, detailed routing or signoff; conclusions concern the evaluated system configurations rather than isolated models or equal-budget comparisons.
Models, engineering priors, permissions and tool environments differ jointly across configurations, and per-output score ranges overlap, so this case does not establish generalization to unseen designs; the scoring scales express engineering preferences, absolute slack is not period-normalized, and incomplete resource logging makes reported core-hours lower bounds.
