Some modern chips ship with small tunable delay elements in their clock network. After manufacturing, each chip's clock timing can be adjusted to compensate for its own random variations, and a slow chip can be pushed into a faster, more valuable speed grade. Published methods for deciding where to put these elements assume that, once a chip is made, its delays are simply known. On a real tester they are not. The tester applies a setting and a clock frequency, and the chip either passes or fails. A failing chip usually locks up and reports nothing. So the information that is missing is exactly the information about the settings that were too aggressive. Statisticians call this missing not at random: whether a value is observed depends on the value itself. If a method throws those failures away, or fits a smooth model through them, it learns a picture of the chip that is biased toward good outcomes.
This note proposes a method for finding each chip's tuning setting with few tester applications. It does not fit a generic smooth model to the pass/fail landscape over discrete settings. Instead it puts uncertainty on the chip's hidden path delays, which do vary smoothly with manufacturing variation. The pass/fail cliff is then computed exactly from the circuit's known timing constraint graph, and every failed test counts as evidence that some constraint was violated. The quantity we want to measure is how much of the profit gain that published tuning methods report under perfect observation survives when a chip can only be probed through censored tests. We also ask whether respecting the structure of the cliff beats the obvious alternative, a Gaussian process with a censoring-aware likelihood.
Core Claim
The idea is to put the smoothness assumption where it holds and keep the discontinuity exact. Surrogate models such as Gaussian processes, which predict an unknown function by assuming nearby inputs give similar outputs, are known to fail at cliffs. They blur physically real boundaries (Cornford et al.), produce spurious oscillation on integer-valued inputs (Garrido-Merchán et al.), and underestimate risk along sharp safe/unsafe margins (Wachi et al.). Post-silicon timing combines all three problems. Buffer settings are discrete. A chip works until a skew margin is breached by a picosecond and then fails completely. That cliff is also where data goes missing. Fixing the kernel does not remove the difficulty, because the cliff is not a property of an unknown function. It follows deterministically from the circuit's constraint graph. The only unknown is the chip's latent delay vector.
To make this concrete, let $d$ be the chip's delays on the edges of a reduced timing constraint graph, $x$ the vector of discrete buffer settings, and $T$ the clock period under test. A test passes when every setup and hold constraint holds, for example $x_i - x_j + d_{ij} \le T$ for every edge $(i,j)$. The per-chip posterior after tests $(x^{(k)}, T^{(k)}, y^{(k)})$ is
$$ p(d \mid \text{tests}) \;\propto\; p(d)\prod_{k} \Big[\, y^{(k)}\,\mathbf{1}\{d \in \mathcal{F}(x^{(k)},T^{(k)})\} + (1-y^{(k)})\,\mathbf{1}\{d \notin \mathcal{F}(x^{(k)},T^{(k)})\}\Big], $$
where $p(d)$ is a design-time statistical timing prior, $y^{(k)}\in\{0,1\}$ records pass or fail, and $\mathcal{F}(x,T)$ is the set of delay vectors that satisfy all constraints at that setting and period. A failed test does not get dropped. It removes a polytope from the delay space. Because the missingness mechanism is known and determined by the value itself, this is the recoverable kind of informative missingness. In the vocabulary of Graphical Models for Inference with Missing Data, which formalizes when a distribution can be reconstructed from data with structured missingness, the indicator $R$ is a known function of the hidden $d$.
Relation to Prior Work
Two papers supply the optimization machinery, and both motivate the gap. Design-Phase Buffer Allocation for Post-Silicon Clock Binning by Iterative Learning places buffers by turning Sobol-sampled virtual chips into integer linear programs (ILPs). Sampling-based Buffer Insertion for Post-Silicon Yield Improvement under Process Variability does the same with 10,000 Monte Carlo chips. Both assume each manufactured chip's delays are fully observed when its buffers are configured. We reuse their sample-to-ILP construction, but we feed it samples from a per-chip posterior conditioned on censored tests rather than from the population prior. That turns their design-time tool into a post-silicon one.
Fast statistical timing analysis for circuits with Post-Silicon Tunable clock buffers reduces the constraint graph by serial and parallel merges. It reports over 1000× speedup over Monte Carlo with under 1% error on the minimum clock period. We use that reduction to make $d$ low-dimensional enough for posterior sampling. Its Gaussian delay assumption constrains us, so we treat the prior's tail shape as an ablation.
Two missing-data papers motivate the approach but are not used directly. Active Learning with Missing Not At Random Outcomes shows that modeling the missingness indicator improves active acquisition, but it does so on tabular classification with an unknown, learned propensity. Our propensity is known from physics, so we do not learn it. Deep Generative Imputation Model for Missing Not At Random Data treats the mask as a parallel modality for imputation. It shares our view that the mask carries signal, but it has no decision loop and no structural constraints.
Practical Bayesian Optimization for Model Fitting with Bayesian Adaptive Direct Search (BADS) is the main hybrid fix for surrogate failure on rugged landscapes. It pairs local Gaussian processes with a direct-search fallback. It assumes every evaluation returns a number and is designed for roughly 3–20 continuous parameters. Our search space is dozens of discrete buffers, and the evaluations that fail return nothing.
Research Program
The system has three parts.
The first is a posterior sampler over reduced-graph delays $d$. The prior comes from design-time statistical timing analysis, optionally refined by on-chip ring-oscillator readings, following the sensor fusion in Transfer Learning for Minimum Operating Voltage Prediction in Advanced Technology Nodes. The pass/fail likelihood above makes the posterior a prior restricted to an intersection of polytopes and polytope complements. We sample it with elliptical slice sampling and hit-and-run moves, checking feasibility through the reduced constraint graph.
The second is a test-selection rule. The next pair of setting and clock period is chosen to maximize the expected reduction in uncertainty about which frequency bin is achievable. This is a combinatorial choice. We solve it greedily over candidate settings proposed by the posterior-sample ILPs, rather than over the full exponential space.
The third is a final configuration step. It solves a chance-constrained ILP over posterior samples: pick the bin and setting that pass on at least a target fraction of samples. The last setting is then verified once on the tester.
A secondary study covers the population level. Chips that fail early screens never reach tuning. Updating the design prior for the next tapeout from only the surviving chips is therefore a selection-biased estimate. We test whether conditioning on the known screen thresholds corrects this.
Baselines
The naive reference points are untuned chips at nominal buffer settings and plain frequency stepping. The weakest acceptable result is that our method beats both on profit at equal tester budget.
The upper bound is the full-observation per-chip ILP from Design-Phase Buffer Allocation for Post-Silicon Clock Binning by Iterative Learning. Given true delays, it reports about 14% average and up to 26% profit improvement, a 4.23% yield increase, buffers at under 1% of flip-flops, and about 0.5% area overhead on ISCAS89 and TAU 2013 with a 45nm model. Sampling-based Buffer Insertion reports yield gains up to 35.97%. On pci_bridge32 it uses 32 buffers for 3,321 flip-flops to raise yield by over 23%. These numbers are a ceiling, not a competitor. Our headline metric is the fraction of that ceiling recovered per tester application.
The strongest practical competitor is Bayesian optimization over discrete buffer settings with a Gaussian process that treats failures as right-censored, Tobit-style, observations. Each failure is modeled as "objective below zero" rather than dropped. This is the obvious way to handle known-threshold censoring, and it attacks the same question without using circuit structure. We found no reported result on Lacuna for censored-likelihood Gaussian process optimization on post-silicon tuning, so we will run it ourselves. It gets the same chips, the same tester budget, the same candidate pool, and the same hyperparameter-tuning budget as our method, with tuning on a held-out set of circuits. We also run BADS adapted to integer rounding, and a greedy one-buffer-at-a-time search driven by pass/fail outcomes.
Two ablations isolate the key ingredients. The first is our structured posterior with failed tests discarded, a complete-case version. It tests whether the missingness handling matters. The second swaps the Gaussian prior for a Student-t prior. This is a cheap, tractable stand-in for the heavy-tailed α-stable surrogates of Loria et al., which motivate the question but whose intractability we avoid. Beating the censored Gaussian process means reaching the same expected profit with significantly fewer tester applications, or higher profit at a fixed budget, across most circuits. A further requirement is that our method produce no more shipped configurations that fail final verification.
Why not just…?
Why not just use a Gaussian process with a censored likelihood? This is the real threat. When the threshold is known, censored regression is standard and may be enough. We expect it to fall short for a structural reason. The objective is a minimum over many path constraints, so its kinks lie along constraint boundaries. These boundaries are not aligned with the coordinate axes, and a stationary kernel over a space of dozens of buffers cannot place them from a handful of fail bits. This matches the oversmoothing failures documented by Wachi et al. and the finding in BADS that global Gaussian process optimization often did no better than random search on rugged targets. If the censored Gaussian process matches us at equal budget, the structural model is unnecessary, and we will report that.
Why not just read delays from on-chip sensors and run the full-observation ILP? Sensors are proxies. Even with 124 silicon-odometer oscillators fused into a transferred network, the 5nm minimum-voltage study reports 3.89 mV error, rising to 4.66 mV without the odometers. That is the right order of magnitude to blur a picosecond-scale cliff. We include a sensor-prior-plus-ILP variant with no further tests. If it matches tested configurations, then the testing loop is not worth its cost.
Why not just guardband or drop the failures? Guardbanding gives up the 14–36% gains reported above. Dropping failures is the complete-case ablation. Missing-data theory predicts that doing so biases the posterior toward feasibility, because failures are exactly the informative observations (Zhang et al.). The direct measurement of that bias here is the shipped-failure rate of the complete-case variant.
Proposed Experiments
The primary testbed follows the cited buffer papers: ISCAS89 circuits including s38584 and TAU 2013 pci_bridge32, a 45nm variation model, and buffer locations fixed by the design-phase method. Virtual chips are drawn by Monte Carlo. Their true delays are hidden behind a simulated tester that returns only pass/fail at a requested setting and period. Some arms also return noisy sensor readings.
We report three things: expected profit across frequency bins, yield, and the rate at which shipped configurations fail verification. All three are measured as functions of the number of tester applications per chip, from 1 to about 30. To stress heavy tails, we also draw chips from skewed delay distributions while keeping a Gaussian prior, which deliberately mis-specifies it. The decisive check is the profit-versus-budget curve against the censored Gaussian process and the complete-case ablation. Our method should reach most of the full-observation ceiling at a budget where the Gaussian process is still exploring. The gap should widen with buffer count.
The secondary, real-silicon check uses the 16nm and 5nm minimum-voltage datasets. We impose screening cutoffs, so that chips whose true value exceeds a threshold lose their later patterns, and then compare transfer models trained on complete cases against models trained with a censoring-aware likelihood. The test is whether the censoring-aware model gets back to the published 3.89 mV error on held-out uncensored chips. Two outcomes would falsify the population-level claim: the complete-case model stays near that error despite the cutoff, or the censoring-aware model fails to close the gap.
References
- Adding Constrained Discontinuities to Gaussian Process Models of Wind Fields
- Dealing with categorical and integer-valued variables in Bayesian Optimization with Gaussian processes
- Safe Exploration and Optimization of Constrained MDPs Using Gaussian Processes
- Graphical Models for Inference with Missing Data
- Design-Phase Buffer Allocation for Post-Silicon Clock Binning by Iterative Learning
- Sampling-based buffer insertion for post-silicon yield improvement under process variability
- Fast statistical timing analysis for circuits with Post-Silicon Tunable clock buffers
- Active Learning with Missing Not At Random Outcomes
- Deep Generative Imputation Model for Missing Not At Random Data
- Practical Bayesian Optimization for Model Fitting with Bayesian Adaptive Direct Search
- Transfer Learning for Minimum Operating Voltage Prediction in Advanced Technology Nodes: Leveraging Legacy Data and Silicon Odometer Sensing
- Posterior Inference on Shallow Infinitely Wide Bayesian Neural Networks under Weights with Unbounded Variance
- Assessing Fairness in the Presence of Missing Data