Floorplanning is an early step in turning a logical circuit into a physical chip: it positions large circuit modules and constrains the space available for subsequent logic placement and wiring. These decisions affect congestion and timing, but their physical consequences emerge through downstream implementation. For AI agents, the task requires translating engineering constraints into spatial decisions whose quality cannot be judged by successful tool execution alone. We present an evaluation framework that connects submitted layouts to a prescribed downstream implementation flow, checks geometric validity, measures physical quality, and reports resource use. On one industrial design whose 99 macros occupy approximately 71% of a fixed L-shaped layout region, we compare our specialized system Nova-SI with general-purpose engineering agents and agents restricted to file access, obtaining complete physical measurements for 39 layout outputs. Under our 0--100 composite quality score, Nova-SI achieves a mean of 57.97 without online feedback from the prescribed power-grid and placement evaluation, compared with 31.17 for the strongest measured general-purpose configuration, GPT-6 Astra. The difference is supported by raw congestion and timing measurements and persists across scoring-sensitivity checks. Layout-level analysis further shows that similar global utilization can conceal substantially different local congestion. Together, the protocol and case study assess agents through the physical consequences of their delivered layouts and provide evidence about system performance on this industrial instance.
Figures & tables
Figure 1: Evaluation from agent deliverables to physical quality. (a) Fixed inputs and three interfaces; only General permits online PG+placement feedback. (b) Macro-geometry checks and power-and-ground (PG) network construction, placement, global-route estimation, and setup timing analysis connect submissions to measurements; signoff is excluded. The layout is Nova-SI output 02, not a common input. Hatching excludes placement; orange outlines mark the largest and smallest macros. (c) Congestion, timing, wirelength, and utilization scores (C/T/L/U, 0–100) for that output; resources are separate. U is 100 times one minus std utilization over macro-excluded area, not the raw utilization percentage. Congestion/std close-ups share a window; colors show maximum positive overflow across layers per 40 μ m bin, yellow–red at 0–12. Top stages, workstations, and timing waveforms are schematic, not native critical paths. Section 3.3 defines scoring.
Outcomes
Resources
Quality ↑
System / model
Harness
n/N
Inv.
Tok. (M)
Core-h
C
T
Qmax
Qmean
Human reference †
—
1/1
—
—
—
94.17
93.51
91.88
91.88
Nova-SI
—
8/8
0
36.13
22.49
34.67
79.68
81.09
57.97
General
GPT-6 Astra [xhigh]
Codex
4/4
0
131.36
30.79
13.87
59.45
64.75
31.17
GPT-6.1 Sol [xhigh]
Codex
4/4
0
286.50
53.81
7.59
39.89
38.37
22.73
Table 1: Quality comparison with resource context. Higher quality is better; invalid geometry and delivery failure score zero.
Figure 2: Per-output congestion, timing, and quality on aligned scales and run offsets. Each panel contains 30 geometry-passing outputs (colored points), nine geometrically invalid outputs, and one delivery failure (red crosses, scored zero rather than zero raw measurements). Black ticks: all-outcome means; n/N : scored/included counts; colors: systems.
Figure 3: Delivered quality and recorded resources paired by outcome. Both panels contain eight Nova outputs and 32 baseline outcomes. (a) Total tokens, paired with the corresponding output. (b) EDA core-hours: external evaluation for Nova and File-only; recorded solving plus evaluation lower bounds for General. Crosses retain resource use of geometrically invalid outputs and the delivery failure, all scored zero. Quality and resources belong to the same included output at each point; this is not an equal-budget comparison or an exhaustive system-compute account.
Figure 4: Nova-SI layouts with the highest (07, Q=81.09 ) and lowest (06, Q=39.01 ) measured scores. These are separate outputs, not successive optimization steps. Each row shows macros, actual std cells, and congestion from the same measurement: maximum positive overflow across layers per 40 μ m bin, on a shared 0–500 scale, linear over 0–1 and logarithmic thereafter. In 07, vertical open regions connect to the lower std region. Output 06 also leaves broad horizontal whitespace, yet has more extensive hotspots within macro-array channels. Whitespace alone does not characterize the local congestion. White denotes die and dashed cyan core; panels are reconstructed from native coordinate and measurement exports. Timing is measured separately in Table 2 , not inferred from the heatmap.
Figure 5: Highest (r01, Q=64.75 ) and lowest (r02, Q=17.19 ) selected General/GPT-6 Astra outputs, using the same extrema rule as Nova in Figure 4 . Left to right: macros, actual std, and congestion. Both figures use a shared 0–500 scale, linear over 0–1 and logarithmic thereafter, with the maximum positive overflow across layers per 40 μ m bin. Global std utilization is 33.16% versus 33.94%, but total overflow is 4,204 versus 204,229. In r02, broader hotspots occur near std strips and entrances between lower horizontal macros. Output r01 is less congested and outperforms six Nova outputs. White denotes die and dashed cyan core; panels are reconstructed from native coordinate and measurement exports.
Figure 6: A common spatial window: x from 2000 to 3100 and y from 600 to 2600 μ m (upper bounds excluded), selected after full-layout inspection. Each pair shows macros with actual std, then maximum positive edge overflow across layers per 20 μ m display bin, on a shared linear 0–7 scale; colors cannot be compared numerically with the full-layout panels. Native-edge midpoints give 525 and 7,869 positive-overflow edges, with totals of 701 and 8,033, for 07 and 06. The latter is only 17.0% of 06’s global overflow, so this window alone cannot explain the overall score difference.
WNS (ns)
TNS (ns)
Violations
Path family
06
07
06
07
06
07
reg-to-reg
−2.041
−0.231
−54.420
−18.772
467
429
reg-to-mem
−2.491
−0.575
−236.917
−4.106
375
33
mem-to-reg
−0.998
−0.309
−20.041
−1.626
82
18
Table 2: Timing severity and breadth for Nova 06 and 07.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Attribute
Value
Original instances
235,727
Functional non-macros / hard macros / physical-only
235,618 / 99 / 10
Functional-macro library references
12
Fixed signal ports
107
Layout-region / core area (mm 2 )
14.123370 / 14.103612
Total functional-macro area (mm 2 )
10.042939
Appendix
Table 3: Input attributes of the fixed instance. Areas come from native polygons or macro dimensions, not optimized post-placement std statistics; die/core denote database layout boundaries.
System / model
ng
Overflow
Peak
GRC (%)
Length (m)
std (%)
Human reference †
1
142
2.00
0.0014
20.31
32.13
Nova-SI
8
17,563
9.50
0.0745
21.89
32.27
General
GPT-6 Astra
4
281,748
23.25
0.9586
28.88
33.56
GPT-6.1 Sol
4
526,552
19.50
1.7251
31.11
33.80
DeepSeek-V4.1-Flash-1017
3
673,784
43.33
1.5112
27.29
33.39
Appendix
Table 4: Mean raw physical quantities for geometry-passing layouts.
reg → reg (ns)
reg → mem (ns)
mem → reg (ns)
System / model
ng
WNS
TNS
WNS
TNS
WNS
TNS
Human reference †
1
− 0.229
− 15.105
− 0.209
− 0.735
− 0.099
− 0.526
Nova-SI
8
− 1.225
− 44.773
− 1.104
− 41.806
− 0.803
− 10.831
General
GPT-6 Astra
4
− 2.352
− 530.426
− 2.240
− 157.122
− 2.325
− 85.630
GPT-6.1 Sol
4
− 8.830
− 3862.730
− 2.608
− 280.384
− 3.326
− 109.591
Appendix
Table 5: Mean worst-scenario setup timing for geometry-passing layouts.
ID
Overflow
Peak
std (%)
C↑
T↑
Q↑
REF
142
2
32.13
94.17
93.51
91.88
01
9,186
9
32.22
31.42
81.66
58.42
02
10,465
12
32.19
28.18
82.76
56.62
03
7,381
10
32.30
35.37
80.33
60.39
04
34,981
20
32.30
16.09
65.28
41.33
05
27,867
12
32.26
20.12
88.62
52.07
Appendix
Table 6: Per-output Nova-SI quality and the historical human reference.
System / model
N
Measured
Inv.
Failed
Nova-SI
8
8
0
0
General
GPT-6 Astra
4
4
0
0
GPT-6.1 Sol
4
4
0
0
DeepSeek-V4.1-Flash-1017
4
3
0
1
GLM-5.3
4
4
0
0
Appendix
Table 7: Physical-measurement and delivery outcomes for the full cohort.
System / model
ng
Qfirst
Qselected
ΔQ
Nova-SI
8
57.97
57.97
+ 0.00
General
GPT-6 Astra
4
31.17
31.17
+ 0.00
GPT-6.1 Sol
4
26.19
22.73
− 3.46
DeepSeek-V4.1-Flash-1017
3
21.58
21.58
+ 0.00
GLM-5.3
4
18.80
18.92
+ 0.12
Appendix
Table 8: Quality at the first complete measurement and the selected layout.
First / selected candidate
n
Recorded execution order
Same
4
One attempt only
Same
10
Attempts before first complete observation
Same
2
First attempt complete; later attempts follow
Different
3
Another candidate selected
Appendix
Table 9: PG+placement attempt ordering in the 19 measured General runs.
System / model
Input tokens (M)
Output tokens (M)
EDA (core-h)
Attempts
Counted/ N
Human reference †
—
—
—
—
—
General
GPT-6 Astra
130.729
0.631
30.79
2.25
4/4
GPT-6.1 Sol
285.621
0.883
53.81
2.75
4/4
DeepSeek-V4.1-Flash-1017
156.154
0.292
25.89
2.00
3/4
GLM-5.3
61.486
0.622
22.23
4.25
4/4
Appendix
Table 10: Mean resources and PG+placement attempts by baseline configuration.
ID
Input tokens (M)
Output tokens (M)
Total tokens (M)
EDA (core-h)
01
20.161
8.349
28.510
17.00
02
22.184
11.247
33.431
19.37
03
19.402
19.866
39.268
19.29
04
18.769
12.818
31.587
26.71
05
9.099
9.000
18.099
23.89
06
21.287
13.790
35.077
21.91
Appendix
Table 11: Per-output Nova-SI resource records. EDA hours come from external evaluation.
Figure 7: Two measured layouts under the same model and harness. Each pair combines macros and actual std at left, with maximum positive edge overflow across layers per 40 μ m bin at right on a shared linear 0–9 scale; colors cannot be compared numerically with the full-layout panels. In r02, region A contains std strips separated by horizontal small macros; several entrances coincide with hotspots absent at the same intensity in r01. Both sets of 99 macros pass geometry checks. Reconstructed from native post-placement exports: this locates differences without isolating a causal effect of channel width.
Figure 8: Low overflow need not indicate a good layout: File-only/GLM-5.3-Flash r04. The axes cover all macros, including those outside the white die boundary. Native overflow is 1, but the geometry audit identifies 53 functional macros outside the die, giving effective score zero. Low overflow coexists with severe geometric violations here; the figure alone does not identify the specific cause of low overflow. The triptych uses a 0–500 color scale.
Coding agents have demonstrated strong capabilities in software development, yet their ability to tackle long-horizon engineering tasks governed by physical constraints remains less understood. We study this question through analog circuit layout, which demands domain knowledge, geometric reasoning, and iterative interaction with design and verification tools. We introduce AnalogLayoutWorld, an open-source agent benchmark with 40 expert-reviewed tasks for end-to-end layout generation from fixed transistor-level netlists. It combines an integrated design environment with independent physical and post-layout electrical verification, as well as objective layout quality metrics co-designed with analog layout experts. Across 13 model--harness configurations, 91.1% of attempts produce GDS layouts, but only 23.0% satisfy the benchmark's acceptance criteria. We also evaluate Novara, an internally developed agent system, which achieves the highest reported aggregate score. Our analysis links performance differences to test-time scaling under fixed budgets, suggesting that reasoning efficiency and effective use of verification feedback help sustain design quality as task complexity grows.
Large language model (LLM) agents are being explored for circuit design, but their ability to produce high-quality custom layouts has not been extensively evaluated. ICLayout-Bench is a benchmark in which agents use electronic design automation (EDA) tools to construct and optimize GDSII layouts from transistor-level circuit netlists. These layouts are evaluated using physical verification, parasitic extraction, and post-layout simulation. Layouts that pass the physical and functional checks are scored on electrical performance and area. The benchmark includes 165 tasks across 13 process design kits (PDKs), covering technology nodes from 3 nm to 180 nm. The tasks span analog, mixed-signal, and digital circuits in open-source and commercial EDA flows. The experiments evaluate 22 configurations spanning seven models on 21 public and 19 private tasks. Mean layout scores range from 4.22 to 80.26 on public tasks and from 1.43 to 67.01 on private tasks, on a 100-point scale. Higher reasoning effort does not consistently improve mean scores, and configurations with comparable scores can incur substantially different model costs.
Zhangmai Li, Tianrui Ma, Zheng Weng +8
University of Science and Technology of China · State Key Lab of Processors, Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences +1
Macro placement still requires substantial manual refinement in industrial physical design flows. We present MAGE (Macro Placement Agentic Engine), a multimodal multi-agent framework for macro placement refinement. MAGE decomposes the macro placement task into a six-phase workflow that combines structured floorplanning rules, visual checks, and iterative refinement. Expert floorplanning knowledge is encoded through natural-language directives and validation criteria, rather than learned from labeled placement data. A tournament-style refinement mode evaluates multiple candidate placements and propagates feedback from higher-quality solutions. We also introduce four metrics for quantifying human-likeness in macro placement: notch score, whitespace score, pocket score, and alignment score. These metrics capture structural properties used by expert designers but not directly measured by conventional PPA metrics. Across nine designs in NanGate45 and GlobalFoundries 12nm enablements, MAGE achieves geometric-mean improvements of 11.1%-19.3% in WNS and 70.0%-74.0% in TNS over commercial macro placers. On the three NanGate45 designs, for which human-expert and Hier-RTLMP baselines are available, MAGE improves WNS and TNS by 18.3% and 72.5% over the human expert, and by 47.0% and 80.4% over Hier-RTLMP, with comparable wirelength and power. On human-likeness metrics, MAGE improves the overall score by 6%-48% over all baselines. Additional case studies on anonymized netlists, unseen designs, dense rectilinear floorplans, and high-utilization settings show that the framework transfers to new placement settings without design-specific retraining.