TAPCO-G: LLM-Guided Design-Adaptive Timing and Power Co-Optimization with Graph-Based Physical Priors
Authors: Wuqian Tang, Zixiao Wang, Che-Chien Lin, Fong-Hua Fu, Laura Angélica Shion-Korsak, Xinyun Zhang, Bei Yu, Chun-Yao Wang
Organizations: National Tsing Hua University, Taiwan · The Hong Kong University of Science and Technology (Guangzhou), China · ShanghaiTech University, China · The Chinese University of Hong Kong, Hong Kong
Post-placement timing and power co-optimization is critical to meeting setup timing targets and reducing power within placement constraints. A cell replacement can improve timing along one path while increasing power or reducing slack on others. Conventional optimizers follow fixed rules for scheduling and accepting repairs, limiting their ability to address design-specific bottlenecks. A large language model (LLM) can propose changes to these rules, but on its own it cannot reliably estimate how a cell replacement affects endpoint slack and power. TAPCO-G combines LLM-guided optimizer adaptation with graph-based physical priors to evaluate candidate cell replacements. The outer loop tunes tool settings using physical-design feedback and revises executable repair policies when tuning stalls. The inner loop anchors delay estimates to static timing analysis snapshots, accounts for upstream driver-load changes, and propagates delay shifts to timing endpoints. It combines these estimates with coarse-to-fine power screening to prioritize candidates for physical trials. TAPCO-G produced placement-legal, logic-equivalent implementations and outscored CERBERUS on all eight ASAP7 benchmarks from the MLCAD 2026 Contest, with a mean quality score gain of 5.516 points, and scored 21.648 points higher than TAPCO on average. These results demonstrate the effectiveness of the proposed co-optimization framework.
Figures & tables
Figure 1. Evolution of post-placement timing and power optimization paradigms across macroscopic policy spaces ( Ωd ) and microscopic timing–power graphs ( Gd=(Vd,Ed) ). (a) Static heuristic flows fix (C0,θ0) and evaluate isolated single-vertex moves on Gd , missing fanin edge load penalties and multi-hop cone conflicts. (b) Black-box parameter tuning explores θ∈Θd(C0) within a single fixed policy plane and stalls at an expressivity ceiling. (c) Uncoupled outer-loop LLM policy adaptation ( He et al., 2023 ; He et al., 2025 ) expands macroscopic policies ( C0→C1 ) but still relies on myopic 1-hop vertex metrics on Gd . (d) Bi-level TAPCO-G jointly evolves outer source policies to πd⋆ and propagates inner replacements across coupled multi-hop DAG cones on Gd . Four-panel schematic comparing static heuristic flow, black-box parameter tuning, outer-only LLM adaptation, and bi-level TAPCO-G across policy planes and timing-power DAGs.
Figure 2. Electrical and physical coupling fundamentals in post-placement co-optimization. (a) NLDM lookup surface Dc(σin,Cload) and physical driver–sink capacitive coupling: upsizing sink i ( c→c′ ) increases pin capacitance Cppin , shifting the upstream driver operating point into a steeper delay/slew regime ( σ↑ ). (b) Instance-level switching activity density ρp governs the balance between dynamic switching power Pdyn(ρp) and multi-VT static leakage Pleak(VT) across RVT, LVT, and SLVT flavors. Two-panel physical foundation showing the nonlinear NLDM delay-slew-load surface and driver-sink coupling on the left, and activity-dependent dynamic versus multi-VT leakage power tradeoff on the right.
Figure 3. Bi-level methodological overview of TAPCO-G . (a) The outer-loop LLM optimizer adaptation engine ( Section 3 ) structures coupled (C,θ) slices, replays evidence-conditioned edits from the clean placement anchor xd0 , and expands source policies when parameter sweeps stagnate. (b) The inner-loop graph-guided replacement engine ( Section 4 ) couples snapshot NLDM differentials with upstream driver loads, propagates clipped arrival shifts across reconvergent cones, and screens candidates via an activity-weighted power funnel and recursive batch bisection. Two-tier three-card methodological overview of TAPCO-G showing the outer-loop LLM adaptation stages on top and the inner-loop graph-guided physical prior stages on bottom.
Figure 4. Hierarchical adaptation space and stagnation-triggered policy escalation. (a) Within a fixed policy slice Ωd(Ck) , scalar parameter tuning stalls at an expressivity ceiling ( ΔSd(m)≤ϵstag over window W ) as static traversal repeatedly hits blocked cones. (b) Detecting stagnation triggers a source-policy edit δCk ( Section 3.4 ), transitioning from Ωd(Ck) to Ωd(Ck+1) to unlock higher QoR frontiers. Comparison between parameter tuning stalling at a policy slice ceiling over window W versus stagnation-triggered source-policy escalation jumping to an expanded policy slice.
Figure 5. Diagnostic-guided LLM proposal and common-anchor execution protocol. (a) Scalar black-box scores mask distinct physical slack distributions while chained netlist executions accumulate placement and electrical drift. (b) TAPCO-G conditions proposals on multi-channel diagnostics gdk and replays every candidate radially from the clean anchor xd0 ( Section 3.3 ), logging all trials in Hdk while monotonically updating the valid incumbent Idk . Comparison showing scalar ambiguity and chained netlist drift on the left, versus multi-channel diagnostic conditioning and radial common-anchor replay with monotonic incumbent tracking on the right.
Figure 6. Circuit-to-graph mapping and snapshot-anchored edge coupling on Gd=(Vd,Ed) . (a) Physical netlist cells, flip-flops, and interconnects map to pin-transition vertices v∈Vd annotated with snapshot state (avs,svs,ρp) and directed cell/net timing arcs e∈Ed . (b) Replacing a candidate vertex sub-graph ( c→c′ ) anchors differential NLDM lookups to τas and couples intrinsic cell-arc speedups with upstream fanin driver-edge capacitance perturbations ΔCp . Two-panel graph formulation showing physical circuit mapping to annotated timing-power DAG G_d on the left, and snapshot-anchored differential cell and upstream fanin edge coupling on the right.
Figure 7. Event-driven topological wave propagation, reconvergent pruning, and endpoint slack clipping on Gd=(Vd,Ed) . (a) Arrival shifts δav(q) propagate across the active fanout cone on Gd while branches masked by dominant reconvergent paths ( ∣δav∣≤ϵt ) are immediately pruned. (b) At sink endpoints Ed(q) , positive slacks are clipped at 0 ns via min(0,se) so non-violating sinks ( e1 ) contribute zero artificial gain while true violating sinks ( e2 ) are accurately priced. DAG wave propagation showing active fanout cone and pruned reconvergent branches on the left, and 0-ns endpoint slack clipping across sink vertices on the right.
Figure 8. Cross-stage sequential graph coupling across boundary vertex vFF=(Dr,Qr) on Gd . (a) Severing the timing graph at flip-flop boundaries isolates upstream cone Gup from downstream cone Gdn and misses capture setup degradation ( δhD ). (b) TAPCO-G couples capture setup shifts δhD on incoming edges to Dr with launch Clock-to-Q arrival waves δaQ on outgoing edges from Qr across vFF . Comparison between severed single-cone DAGs at flip-flop boundaries versus coupled bi-stage DAG cones propagating both capture setup shifts and launch arrival waves across sequential vertices.
Figure 9. Graph-wide coarse-to-fine node power screening and conflict-free batch bisection on Gd . (a) TAPCO-G prunes low-leverage vertex families via Fc and refines active graph vertices using pin switching density ρp and slack-risk priority R(q) . (b) Top-ranked vertices with disjoint fanout cones are grouped into conflict-free batches B⊂Vd , while any batch triggering nonlinear regression is recursively bisected ( B→BL,BR , Section 4.5 ) to isolate conflicting vertices. Two-panel graph view showing coarse-to-fine activity-weighted vertex screening on G_d on the left, and independent vertex batching with recursive bisection conflict isolation on the right.
Design
Instances
TNS (ns)
Total (mW)
Leakage ( μ W)
SRAM
AES
20,097
-12.44
432.0
96.3
No
JPEG
64,602
-48.77
294.0
174.0
No
NVDLA-P
106,994
-198.18
40.0
306.0
No
Ariane
210,455
-7568.43
658.0
17900.0
Yes
NVDLA-A
75,686
-203.23
165.0
321.0
No
NVDLA-C
312,696
-55.58
601.0
17300.0
Yes
Table 1. Initial benchmark characteristics of the eight-design ASAP7 suite under the reference contest constraints. Instance counts include all physically instantiated standard cells and hard macros.
Design
OR-LR ( 2026 )
OR-LR+PR ( 2026 )
TAPCO ( 2026 )
CERBERUS ( 2026 )
TAPCO-G
Δ TAPCO
Δ CERB.
AES
1.354
−18.360
37.778
58.502
64.729
+26.951
+6.227
JPEG
29.951
−32.809
29.801
42.991
56.928
+27.127
+13.937
NVDLA-P
51.261
−121.743
Fail
59.341
63.325
Valid ( ✓ )
+3.984
Ariane
0.102
−0.290
17.074
27.549
27.870
+10.796
+0.322
NVDLA-A
42.620
−7.892
Fail
58.779
71.028
Valid ( ✓ )
+12.249
NVDLA-C
25.071
−1607.154
Fail
27.844
28.754
Valid ( ✓ )
+0.910
Table 2. Main quality score ( SQ ) comparison across all eight ASAP7 benchmarks. OR-LR+PR reports the direct post-power-recovery score; because uncoupled power recovery degrades every design, a best-of-stage checkpoint selector between OR-LR and OR-LR+PR simply reverts to OR-LR on all eight benchmarks (mean 21.164 ). Fail indicates that the method failed to produce a placement-legal, LEC-verified final-phase implementation. Bold highlights the best result on each design.
Final Physical Signoff Metrics ( TAPCO-G )
Score Decomposition (pts)
Design
WNS (ns)
Final TNS (ns)
Total Power (mW)
Leakage ( μ W)
Timing
Dyn.
Leak.
Pen.
SQ
AES
−0.096
−12.15 ( −2.3% )
259.0 ( −40.0% )
11.5 ( −88.1% )
+0.70
+20.02
+44.03
−0.02
64.73
JPEG
−0.102
−38.34 ( −21.4% )
225.0 ( −23.5% )
38.9 ( −77.6% )
+6.42
+11.72
+38.82
−0.03
56.93
NVDLA-P
−0.351
−69.05 ( −65.2% )
38.4 ( −4.0% )
48.3 ( −84.2% )
+19.55
+1.69
+42.11
−0.02
63.33
Ariane
−0.256
−602.92 ( −92.0% )
658.0 ( 0.0% )
17800.0 ( −0.6% )
+27.61
−0.01
+0.28
−0.01
27.87
NVDLA-A
−0.188
−118.98 ( −41.5% )
111.0 ( −32.7% )
47.3 ( −85.3% )
+12.44
+16.31
+42.63
−0.35
71.03
Table 3. Detailed physical signoff metrics and decomposed score components of TAPCO-G across all eight benchmarks. Relative percentages in parentheses indicate reductions relative to the initial placed baselines in Table 1 . Following Eq. 8 , Timing+Dyn.+Leak.+Pen.=SQ , where Pen.=−∑jλjpd,j combines the displacement penalty −0.5Davg (mean Davg=0.044 across the suite) and residual electrical slew/capacitance penalties (e.g., −0.34 electrical and −0.01 displacement points on NVDLA-A ).
Figure 10. Decomposition of TAPCO-G signoff quality scores into normalized setup timing ( ΔTd ), dynamic power ( ΔPdyn ), and leakage power ( ΔPleak ) contributions across all eight benchmarks. Diamonds mark the complete signoff score SQ after electrical, displacement, and routing adjustments. Horizontal bar chart decomposing TAPCO-G scores across eight benchmarks into timing, dynamic power, and leakage gains.
Initial Baseline
TAPCO ( 2026 )
TAPCO-G (Ours)
Design
TNS
Pdyn
Pleak
TNS
Pdyn
Pleak
Score
TNS
Pdyn
Pleak
Score ( Δ )
AES
−12.44
431.9
96.3
−17.67
385.0
9.7
37.778
−12.15
259.0
11.5
64.729 ( +26.951 )
JPEG
−48.77
293.8
174.0
−68.51
288.0
31.5
29.801
−38.34
225.0
38.9
56.928 ( +27.127 )
Ariane
−7568.43
640.1
17900.0
−3092.71
645.0
18000.0
17.074
−602.92
640.2
17800.0
27.870 ( +10.796 )
NVDLA-M
−13.99
44.5
61.5
−22.49
43.9
9.5
24.737
−15.89
38.5
7.6
46.455 ( +21.718 )
Table 4. Head-to-head physical PPA comparison between TAPCO ( Tang et al., 2026 ) and TAPCO-G on the four designs where TAPCO produced valid final-phase solutions (out of eight total benchmarks). Units: TNS in ns , dynamic power Pdyn in mW , and leakage power Pleak in μW .
Figure 11. Physical comparison between TAPCO-G and TAPCO . (a) Component-wise timing ( ΔTd ), dynamic-power ( ΔPdyn ), and leakage ( ΔPleak ) score gains of TAPCO-G over TAPCO under identical (30,50,50) weights. (b) Standard-cell VT (RVT, LVT, SLVT) distributions in the emitted implementations. Left panel shows stacked bar improvements of TAPCO-G over conference TAPCO in timing and dynamic power; right panel shows multi-VT cell population breakdowns.
Figure 12. Predicted versus physically measured singleton timing–power gain across 77 independent cell-replacement probes ( 45 on AES and 32 on NVDLA-M ). Each point is physically applied, evaluated via full signoff STA and power analysis, and exactly restored to the baseline state. The dashed diagonal ( y=x ) marks ideal zero-error prediction. Scatter plots on AES and NVDLA-M showing that graph-guided predictions closely track measured physical gains along the y=x diagonal whereas local predictions exhibit large scatter and sign errors.
Design
Timing–Power Predictor
MAE (pts) ↓
Sign Accuracy ↑
Top-8 Hit Rate ↑
AES
Myopic Local Estimator
0.0683
34/45 ( 75.6% )
5/8 ( 62.5% )
AES
Graph Propagation (No Load)
0.0183
40/45 ( 88.9% )
7/8 ( 87.5% )
AES
Full Graph Model ( TAPCO-G )
0.0185
42/45 ( 93.3% )
8/8 ( 100.0% )
NVDLA-M
Myopic Local Estimator
0.1283
26/32 ( 81.3% )
6/8 ( 75.0% )
NVDLA-M
Graph Propagation (No Load)
0.0175
29/32 ( 90.6% )
7/8 ( 87.5% )
NVDLA-M
Full Graph Model ( TAPCO-G )
0.0364
31/32 ( 96.9% )
8/8 ( 100.0% )
Table 5. Singleton prediction fidelity across 77 independent physical replacement probes on AES ( 45 probes) and NVDLA-M ( 32 probes), comparing myopic local estimation, uncoupled graph propagation, and the full TAPCO-G graph model with driver-load coupling.
Ablated Component
Design
Budget
Accepted (OFF → ON)
Score OFF
Score ON
Gain ( ΔSQ )
Graph Prop. (COMB)
AES
256
70→185 ( 2.64× )
60.201
60.821
+0.620
Graph Prop. (COMB)
NVDLA-M
256
39→188 ( 4.82× )
42.284
43.448
+1.164
Graph Prop. (FF+COMB)
AES
128
64→112 ( 1.75× )
60.293
60.635
+0.341
Graph Prop. (FF+COMB)
NVDLA-M
128
51→104 ( 2.04× )
42.236
42.489
+0.253
Graph Prop. (Late-Stage)
AES
256
82→164 ( 2.00× )
61.848
62.164
+0.316
Graph Prop. (Late-Stage)
NVDLA-M
256
48→142 ( 2.96× )
45.038
45.264
+0.226
Table 6. Matched component ablations of the inner-loop physical priors under identical initial checkpoints, binaries, and root-action budgets. Each row isolates the effect of enabling a single physical-guidance mechanism.
Figure 13. Monotonic signoff quality ( SQ ) progression across retained adaptation stages on all eight ASAP7 benchmarks, illustrating how TAPCO-G combines outer-loop policy/parameter specialization with inner-loop graph-guided continuations to reach design-specific optima. Line plots showing monotonic score improvements across retained adaptation stages for all eight benchmarks.
Design
Wire ( μ m)
Slew/Cap. (%)
Setup Margin (ns)
Repairs/Pass
Primary Active Move Classes
AES
60
10/10
+0.500
50
Selective Sizing + Graph Multi-VT
JPEG
Default
Default
−0.105
Default
Graph Multi-VT + Load-Coupled Sizing
NVDLA-P
150
19/19
Default
2
Conservative Sizing + Power Funnel
Ariane
70
Default
+0.210
200
Full Buffering, Cloning, Sizing & VT
NVDLA-A
78
9/9
+0.050
2
Tight-Envelope Sizing + Graph Multi-VT
NVDLA-C
30
20/20
+0.032
25
Targeted Critical-Cone VT & Pin Swap
Table 7. Representative design-specialized parameter configurations ( θd⋆ ) discovered by TAPCO-G across the eight benchmarks. Default denotes the inherited baseline setting of the underlying repair pipeline.
Policy Dimension
Expressivity Limit of Fixed Parameters θ
Synthesized Source-Level Policy Cd⋆
Dynamic Frontier ( Dsched )
Static endpoint traversal order ignores timing shifts induced by earlier moves.
Dynamically recomputes and re-sorts the violating endpoint frontier after intermediate repair passes.
Severity Effort ( Deffort )
Uniform pass budgets waste whitespace on shallow paths or hammer blocked cones.
Allocates transformation effort proportionally to path violation depth and bypasses saturated cones.
Analytical placers rely on differentiable objective functions to guide placement, typically combining intermediate surrogate metrics such as half-perimeter wirelength (HPWL) and cell-density penalties. However, these placement-stage surrogates remain misaligned with downstream routed and timing quality. Prior work reduces this gap with human-designed terms or learned black-box surrogates, but the former requires expert retuning and the latter is difficult to explain, debug, or deploy in analytical placement flows. CoEvoP&R addresses these limitations with a large language model (LLM)-based framework that automatically evolves analytical placement objectives. At each generation, the prompt combines the restricted objective interface, baseline context, and archived prior candidates with routing-related feedback from placement, timing proxy, and routing tools. The LLM proposes readable differentiable objectives, which are embedded and validated in DREAMPlace, evaluated through a timing proxy and an actual router, and stored with their feedback to guide later generations. Across eight ChiP-Bench Nangate45 designs and three seeds, CoEvoP&R reduces post-route routed wirelength and congestion by 16.9% and 36.7%, with gains of 0.70 ns in worst negative slack and a 912 ns reduction in total negative slack magnitude over native DREAMPlace. Across eight ICCAD 2015 Superblue designs, it reduces post-route routed wirelength and congestion by 5.4% and 23.2%. Code is available at https://github.com/FCHXWH823/CoEvoP-R.git.
Ruogu Chen, Weihua Xiao, Ramesh Karri +1
University of Alberta Edmonton, Alberta, Canada · New York University New York, New York, United States
Applying large language models (LLMs) to RTL code optimization for improved power, performance, and area (PPA) faces two key challenges: ensuring functional correctness of optimized designs despite LLM hallucination, and systematically prioritizing power reduction within the multi-objective PPA trade-off space. We propose POET (Power-Oriented Evolutionary Tuning), a framework that addresses both challenges. For functional correctness, POET introduces a differential-testing-based testbench generation pipeline that treats the original design as a functional oracle, using deterministic simulation to produce golden references and eliminating LLM hallucination from the verification process. For PPA optimization, POET employs an LLM-driven evolutionary mechanism with non-dominated sorting, power-first intra-level ranking, and proportional survivor selection to steer the search toward the low-power region of the Pareto front without manual weight tuning. Evaluated on the RTL-OPT benchmark across 40 diverse RTL designs, POET achieves 100% functional correctness, the best power on all 40 designs, and competitive area and delay improvements.
Post-route power recovery changes cell choices under tightly coupled timing, electrical and physical constraints, and the value of each change depends on what it costs to repair its side effects. We propose an AI-driven framework in which a learned model drives both phases of the optimization: it ranks the powerrecovery edits to try and the repair actions that follow, while deterministic electronic design automation (EDA) analyses govern every acceptance. The framework records baseline constraint margins, identifies the damage a recovery transaction introduces in each analysis view, holds pre-existing violations at their baseline rather than working them inside the power step, and caps repair spending at a fraction of the measured saving. An ensemble of gradient-boosted models predicts power change, normalized damage reduction and constraint-regression risk from action and timing-context features available before each trial; calibrated, conservative estimates feed separate acquisition scores for recovery and repair, with a deterministic fallback when the model is uncertain. Recovery and repair remain speculative until physical implementation and fresh analysis establish baseline preservation and positive net saving. We specify the contract, the learned policy, the transactional flow and a pre-registered public evaluation protocol on open designs. Evidence from a companion study on public designs supports the contract; the learned policy is proposed, and no trained-model results or proprietary design measurements are reported.