Chip placement tools decide where large memory blocks ("macros") sit on the silicon. They are judged by what happens much later, when wires are actually routed and timing is checked. Those downstream checks take minutes to hours per layout, so placers rely on cheap stand-ins. One is wirelength. Another is a hand-written rule that pushes macros toward the chip's edges to keep the center open for wiring. A third is a learned predictor trained on past routing runs. The natural next step is to run the real downstream check on some candidate layouts during the search, in parallel on a cluster, and feed the results back. This creates a problem placement research has not named. Real routing results come back late, and they describe layouts the search has already moved past. The search ends up steering by stale information.
We propose treating this as an Age of Information problem. Age of Information is a measure, from networking, of how long ago the newest piece of information a decision-maker holds was generated. We would build a freshness-aware evaluation dispatcher that decides which layouts to send for real routing, drops queued jobs that have been superseded, and controls how much the search leans on the periphery rule. That reliance grows with how stale the learned predictor's information is. The dispatcher sits inside Parsl, a Python library that runs the same script on a laptop or a cluster. A placement researcher would get distributed, freshness-managed downstream evaluation without writing cluster code.
Core Claim
The claim has two parts, which can be tested together.
First, under a fixed compute budget, asynchronous downstream evaluation with a first-in-first-out job queue wastes much of its budget on stale labels. Queueing theory predicts this. Optimizing Age-of-Information in a Multi-class Queueing System shows that expected peak age equals the mean time between updates plus the mean time an update spends in the system. When updates arrive faster than the server can handle them, queueing delay dominates and age "blows up." The same paper shows that a queue that discards arrivals when the server is busy avoids this. In our setting the "server" is the pool of routing workers, and an "update" is a routed-and-timed label for a candidate layout.
Second, the periphery bias is most valuable exactly when the learned predictor's information is stale. It is a robust prior that does not depend on recent labels. It should therefore not have a fixed weight. Its weight should rise as labels age.
Concretely, let $g(t)$ be the search generation at wall-clock time $t$. Let $u(t)$ be the generation of the newest layout whose downstream label has been incorporated into the predictor by time $t$. The predictor's age is
$$A(t) = g(t) - u(t).$$
Age is measured in search steps rather than seconds because what matters is how far the search has moved since the newest label. The placer scores a candidate layout $p$ by
$$J(p) = \big(1 - w(A)\big)\, \hat{c}(p) + w(A)\, r(p),$$
where $\hat{c}(p)$ is the learned predictor's estimate of downstream cost (congestion and negative slack), $r(p)$ is the periphery-regularity penalty, and $w(A) \in [0,1]$ is an increasing function of age that we fit on held-out designs.
Research Program
The search side starts from existing methods. LaMPlace trains a graph neural network offline on 1,200 DREAMPlace layouts labeled by a commercial flow. It predicts downstream timing and congestion as a cost map over the canvas. LaMPlace observes that edge-hugging macro placement emerges from this learned cost without being hard-coded. The research file lists offline-to-online distribution shift as its main limitation, because the predictor never sees labels from the region the online search explores. Reinforcement Learning Policy as Macro Regulator Rather than Macro Placer, called MaskRegulate, refines a DREAMPlace layout with a reinforcement-learning agent. Its reward includes a fixed-weight regularity term measuring each macro's distance to the nearest boundary. Macro Placement by Wire-Mask-Guided Black-Box Optimization, called WireMask-EA, runs an evolutionary search over unconstrained coordinates. A greedy wire-mask legalizer turns each candidate into a legal layout.
We use WireMask-EA as the search engine for three reasons: it is fast, it is population-based, and it parallelizes naturally. The LaMPlace-style predictor serves as $\hat{c}$, updated online by fine-tuning on fresh labels. MaskRegulate's regularity mask serves as $r$. Neither prior system fetches downstream labels during search, and neither adapts the regularity weight. Those are the two things we add.
The dispatcher borrows its policies from the Age of Information literature, each mapped to a concrete placement decision.
The queue discipline follows two papers. Optimizing Age-of-Information in a Multi-class Queueing System motivates discarding: a pending routing job is cancelled once a newer candidate from the same lineage exists. Age of Information: The Gamma Awakening shows that serving the newest request first without interrupting the current job beats interrupting it when service times are near-deterministic and requests arrive often. Routing jobs on a fixed design have fairly stable runtimes, so we test both the interrupting and non-interrupting variants rather than assuming interruption is best.
The decision of when to request a label follows Joint Status Sampling and Updating for Minimizing Age of Information in the Internet of Things. That paper proves that sending an update immediately after the last one finishes is suboptimal, and that the optimal policy has an "idle region" where the device waits because the receiver's information is still fresh enough. In placement terms, the dispatcher skips evaluation when $A(t)$ is below a threshold.
Efficient Information Updates in Compute-First Networking suppresses status updates when the monitored quantity has not changed, combining age with a value term. We adopt the analogous idea: a label is requested only if the candidate's predicted cost differs from its nearest labeled neighbor's by more than the predictor's recent error.
Age-Based Scheduling for Mobile Edge Computing redefines age for tasks that need computation before their result is usable. We follow it and count a label as delivered only when routing and predictor fine-tuning have both finished.
The infrastructure side is built on Parsl. Parsl wraps Python functions and command-line tools as asynchronous tasks that return placeholders ("futures"). It builds the dependency graph at runtime and hides whether the work runs locally, under Slurm, or in a cloud. The research file notes that Parsl and similar workflow engines schedule purely by dependency readiness and have no notion of freshness. Our dispatcher adds that notion as a scheduling layer: tasks carry a generation stamp, and superseded futures are cancelled.
The deliverable is a single Python entry point. The downstream flow is pluggable: an open-source global router plus DREAMPlace cell placement for anyone, and Cadence Innovus where a license exists. This lowers the barrier described by the research file, where distributed placement exploration currently requires hand-written shell scripts and commercial-tool scripting.
Baselines
The naive reference point is WireMask-EA optimizing wirelength alone, with no downstream labels and no regularity term. This is the centering-biased behavior that the periphery-placement literature argues produces unroutable layouts.
The fixed-prior baseline is the same search with MaskRegulate's regularity term at a tuned constant weight. It isolates how much the periphery heuristic alone delivers. On superblue1 of the ICCAD 2015 benchmark, MaskRegulate reports a 73.08% reduction in horizontal routing congestion, a 38.81% reduction in vertical congestion, a 17.08% improvement in routed wirelength, and an 18.35% improvement in worst negative slack. Those figures are from Reinforcement Learning Policy as Macro Regulator Rather than Macro Placer, measured against its own baselines.
The strongest existing method is LaMPlace's offline-trained predictor guiding the search, with no online labels. LaMPlace reports average improvements of 43.0% in worst negative slack and 30.4% in total negative slack over the best baselines it compared, as stated in LaMPlace. We will retrain the predictor on our own 1,200-sample offline set so that every method sees the same offline data. We will report our reproduced numbers alongside the published ones, since our downstream flow may differ from theirs.
The online-but-naive baseline sends every new elite candidate to downstream evaluation through a plain first-in-first-out Parsl queue. Labels are fine-tuned into the predictor as they arrive.
The fair-comparison protocol is the same for all methods. They share the offline data, the same number of worker nodes, the same wall-clock budget, and the same cap on total downstream evaluations. Each method gets the same hyperparameter-search budget on held-out designs. Beating the strongest competitor means statistically better final post-route worst negative slack, total negative slack, and congestion than offline LaMPlace and the first-in-first-out online baseline, over at least five seeds, at equal wall-clock time.
Ablations remove one ingredient at a time:
- discarding superseded jobs,
- the idle-region threshold,
- the value-based suppression,
- the age-dependent weight $w(A)$, replaced by its best constant.
Why not just…?
Why not just use the periphery rule at a fixed weight? It may capture most of the gain. The published MaskRegulate numbers are large, and LaMPlace finds edge placement emerging on its own. The fixed-prior baseline and the constant-weight ablation measure this directly. If adaptive weighting adds nothing beyond the best constant, the second half of our claim fails, and we would report the dispatcher as the only contribution.
Why not just use batch Bayesian optimization or uncertainty-based active learning to choose which layouts to label? This is the most serious threat. Uncertainty sampling also concentrates labels where the predictor is unsure. It does not, however, model the delay between request and delivery, so it still queues requests that are stale by the time they finish. We will run an uncertainty-triggered dispatcher both with first-in-first-out queueing and with our discard policy. If uncertainty plus discarding matches the full age-based method, the honest conclusion is that the contribution reduces to the queue discipline. We would report it that way.
Why not just run synchronous batches so labels are never stale? Synchronous rounds remove staleness but leave workers idle while the slowest routing job finishes. They also halt the search during evaluation. We include a synchronous-batch baseline at equal wall-clock time. The test is whether its lost throughput costs more than the staleness it avoids.
Why not just collect more offline data for LaMPlace? Offline data cannot cover the region the online search reaches on a new netlist, which is the distribution-shift limitation named in the research file. We will give offline LaMPlace the full online label budget as extra offline samples, collected before search, to check whether label timing matters beyond label count.
Proposed Experiments
Training and tuning use ICCAD 2015 superblue designs, with the final evaluation on held-out superblue circuits. We then check transfer to ChiPBench industrial designs, which LaMPlace used to test generalization. Wirelength-only sanity checks use ISPD 2005 adaptec designs, where WireMask-EA reports the lowest wirelength on six of seven benchmarks and matches MaskPlace quality in about eight minutes. MaskPlace provides that reference point.
The metrics are:
- post-route worst negative slack, total negative slack, horizontal and vertical congestion, and routed wirelength;
- wall-clock time and the number of downstream evaluations consumed;
- two freshness metrics: mean and peak predictor age $A(t)$, and the fraction of completed evaluations whose layouts were already superseded on arrival.
No reported result was found in the research file for the age or staleness of downstream labels in any placement method. These freshness measurements will be new.
Three checks decide the claims:
- Queueing prediction. As we sweep the request rate, the first-in-first-out dispatcher's peak age should rise sharply once requests exceed worker capacity, as the peak-age decomposition predicts. Final slack should fall with it. The discarding dispatcher should not show this blow-up.
- Age-dependent weighting. The fitted $w(A)$ should beat the best constant weight specifically on held-out and ChiPBench designs, where the predictor starts out of distribution.
- The serious alternative. Our dispatcher must beat uncertainty-plus-discarding. If it only ties, the paper's claim narrows to the queue discipline alone.
References
- Optimizing age-of-information in a multi-class queueing system
- LaMPlace: Learning to Optimize Cross-Stage Metrics in Macro Placement
- Reinforcement Learning Policy as Macro Regulator Rather than Macro Placer
- Macro Placement by Wire-Mask-Guided Black-Box Optimization
- Age of information: The gamma awakening
- Joint Status Sampling and Updating for Minimizing Age of Information in the Internet of Things
- Efficient Information Updates in Compute-First Networking via Reinforcement Learning with Joint AoI and VoI
- Age-Based Scheduling for Mobile Edge Computing: A Deep Reinforcement Learning Approach
- Parsl: Scalable Parallel Scripting in Python
- MaskPlace: Fast Chip Placement via Reinforced Visual Representation Learning