When Does Slow Energy Memory Break Causal Discovery? A Test on Stacked-Chip Heat Flow

Causal discovery methods try to recover which variables directly drive which others from recorded time series. They are hard to evaluate because real datasets rarely come with a known answer. Stacked three-dimensional chips offer an unusual exception. Heat in these chips flows vertically through several tiers of silicon and poorly conducting bonding layers, and a simulator built from the floorplan knows the exact answer: which block heats which, and with what delay. We propose to use this physics as a testbed where the true answer is known.

We then add the ingredient that makes causal discovery genuinely hard. That ingredient is a hidden common driver with long memory. On an energy-harvesting device, the power drawn by every block rises and falls together with the harvested energy and the battery policy. That energy is temporally correlated: it stays roughly constant for a "coherence time" and then jumps. The same holds for unmeasured coolant or ambient temperature. The proposal asks one sharp question. How does the memory of this hidden driver, relative to the chip's own thermal time constant, decide whether standard methods invent false heat-flow edges and whether latent-variable methods are actually needed?

Core Claim

Our hypothesis concerns a ratio. Call $T$ the correlation time of the hidden driver and $\tau$ the thermal time constant of a block, meaning how slowly its temperature responds to a power step. When $T \ll \tau$, heat diffusion averages the driver into something close to white noise, and ordinary lagged causal discovery should work. When $T \gtrsim \tau$, the driver leaves a slow shared trend in every tier, and ordinary methods will report dense, false vertical edges.

The expected form of this transition comes from filtering. If a first-order thermal filter with time constant $\tau$ receives an input with exponential autocorrelation of time scale $T$, the fraction of the input's variance that survives is $T/(T+\tau)$. The spurious shared signal should therefore grow roughly like that ratio. We predict that the precision of methods that ignore hidden drivers falls along this curve, while a latent-variable model with a prior matched to the driver's memory stays flat.

The deliverable has three parts. The first is a benchmark with ground truth, a dial for $T/\tau$, and realistic nonlinearities. The second is a measured phase diagram of when each method family fails. The third is a modest method change: a hidden-driver prior shaped like block-constant energy arrivals. We then test whether that change matters beyond simply modeling a hidden autoregressive process.

Grounding in Prior Work

Three groups of papers play different roles in this proposal.

The first group motivates the regime. Energy-harvesting theory shows that arrival memory changes system behavior qualitatively. Online Power Control for Block i.i.d. Energy Harvesting Channels models arrivals as constant within blocks of length $T$. It proves that throughput falls as $T$ grows and that the battery must scale linearly with $T$. Zdeborova et al. show that ignoring correlation badly overstates reliability. These works treat spent energy as vanishing harmlessly; we follow that energy into heat. We reuse the block-i.i.d. arrival model and its fixed-fraction battery policy as the generator for our hidden driver.

The second group supplies the physics and constrains which tools we can use. Cool-3D chains an architecture simulator, power models, and the HotSpot 7.0 thermal network for multi-tier chips on Splash2 workloads. It treats power as a fixed open-loop trace. We instead drive it with correlated, feedback-coupled power. ATSim3D shows two things: ignoring temperature-dependent leakage and conductivity causes errors of up to 10°C in stacks connected by through-silicon vias, and the tool is steady-state only. We therefore model leakage feedback in the transient simulator ourselves, and we use ATSim3D's monolithic and via-based structures only as geometry templates. Open3DBench reports a 10.04% average peak-temperature increase from stacking, and supplies real floorplans such as the Ariane RISC-V core. 2D-ThermAl is a fast forward predictor restricted to planar chips. We do not rely on it, because a forward predictor answers "what temperature" and not "which block causes it."

The third group supplies the methods under test. TL-VSCM models unmeasured interference as a hidden vector-autoregressive process. It learns a binary adjacency matrix through a differentiable relaxation, and reports precision of 0.94–1.0 on fMRI, where DYNOTEARS, VAR-LiNGAM, and PCMCI fall to 0.2–0.4. It is linear, and it has never been tested where the confounder's memory is varied systematically. That gap is the one we fill. SPACY learns spatial compression and a causal graph over hidden factors jointly, and argues that compressing first and discovering second loses weak causal drivers. Its isotropic spatial kernels were designed for climate maps, not layered media. CausalProfiler generates synthetic causal models and exposes brittle failures, such as VACA failing numerically 53.4% of the time in one setting. Our testbed complements it: CausalProfiler's graphs are abstract, while ours come from physics with a dialable confounder memory.

Research Program

Generator. We build the testbed on the public Cool-3D toolchain. We take Open3DBench and Cool-3D floorplans with two to seven tiers, and partition them into 20–200 functional blocks with one temperature sensor each.

Driver and power. Harvested energy follows the block-i.i.d. process with coherence $T$. A fixed-fraction battery policy converts it into a power budget, which is split across blocks according to Splash2 activity traces. A second hidden driver, coolant inlet temperature, follows a slow autoregressive process with its own time scale. Temperature feeds back into leakage power through an exponential leakage law, so a real temperature-to-power loop exists.

Ground truth. The true lagged graph has two kinds of edges. The first are nonzero thermal conductances between blocks, extracted from the discretized resistor-capacitor network. The second are edges from each block's power to its own temperature, plus the leakage return edges. Hidden drivers are withheld from the learner.

Sweep. We vary $T/\tau$ over three orders of magnitude. We also vary which variables are observed: temperatures only, or temperatures plus per-domain power counters. Finally, we vary stacking order, including the memory-buried-under-logic configuration that Cool-3D shows runs hottest.

Method change. We modify TL-VSCM's hidden process in two ways. First, it becomes a switching, piecewise-constant process with learned switch points, matching block-i.i.d. arrivals. Second, the linear mechanism is replaced by a small per-edge nonlinearity to capture leakage. This is deliberately a minimal change: the question is whether matching the hidden prior to the energy statistics matters, not whether a larger model helps.

Baselines

Do-nothing reference. Every pair of physically adjacent blocks is declared connected, with no learning. This shows how much any method adds beyond reading the floorplan. When the floorplan is withheld, the reference becomes a lagged-correlation threshold.

Standard lagged methods. PCMCI and DYNOTEARS, the methods that collapsed in TL-VSCM's experiments. They receive identical series, the same lag window, and the same hyperparameter search budget, tuned on a held-out floorplan rather than the test one.

Strongest existing competitor. TL-VSCM itself, run unchanged, because it is the only method in the file that explicitly models a hidden autoregressive interferer. The strongest reported number is its fMRI precision of 0.94–1.0 against 0.2–0.4 for the classic baselines. We will report whether that gap reproduces on thermal data, and at which $T/\tau$ it opens. Beating it means higher edge precision at matched recall, with lower structural Hamming distance, in the regime $T/\tau > 1$, across at least three held-out floorplans.

Joint spatial model. SPACY applied to the sensor grid. For SPACY on synthetic grids the file gives no per-baseline numbers: "no reported result found on Lacuna for SPACY's F1 values."

Ablation isolating the key ingredient. Our model with the switching hidden prior replaced by TL-VSCM's plain autoregressive prior, all else fixed. A second ablation removes the nonlinear leakage edge.

Upper bound. Every method is also run with the hidden drivers revealed. This measures how much of the failure is due to confounding alone.

Why not just…?

Why not just fit a grey-box resistor-capacitor model with a Kalman filter? When the floorplan and material constants are known, this is a real threat. It will probably match or beat any discovery method, because it has the correct structure built in. We therefore include it as a baseline wherever the floorplan is given, and we expect to lose to it there. The proposal's claim is limited to two cases. The first is a withheld floorplan, as in post-silicon telemetry. The second is bonding-layer conductivities perturbed by ±50%. In both, a misspecified grey-box model absorbs the hidden driver's shared trend as spurious conductances, just as a linear model without latent variables does. If the grey-box model stays accurate under perturbation, the practical case for discovery on chips shrinks to the benchmark contribution alone, and we will say so.

Why not just difference or prewhiten the series to remove the slow driver? Against a block-constant driver this is a serious alternative, since differencing cancels it everywhere except at switch points. It also amplifies sensor noise and damages the slow thermal-lag signal that carries the true vertical edges. We run PCMCI on differenced and on prewhitened series as explicit baselines across the full $T/\tau$ sweep. If differencing alone recovers precision at large $T/\tau$, the switching-prior contribution is unnecessary, and the paper becomes a negative result about latent models. That outcome is still worth reporting.

Why not just measure the harvested power and include it as a covariate? Harvested input is often measurable, and conditioning on it removes that confounder. The revealed-driver upper bound covers this case directly. The coolant or ambient driver remains unmeasured in most deployments, and per-block power split by the scheduler is rarely logged. The experiment reports how much precision each unobserved driver costs separately.

Why not just compress with principal components and then run PCMCI? SPACY argues that variance-based compression discards weak causal drivers. In a stack this matters concretely: buried memory tiers have small temperature swings but, per Hadidi et al., fail at 75°C. We include the two-stage pipeline and score recall on edges into memory tiers specifically.

Proposed Experiments

The primary result is a phase diagram. It plots edge precision and structural Hamming distance against $T/\tau$ for each method, on held-out floorplans. We fit the predicted $T/\tau$-shaped curve to the degradation of the non-latent methods. The first decisive check asks whether that degradation follows the filtering prediction or begins at a different ratio. A mismatch would mean something other than confounding, such as the leakage feedback loop, drives the failures.

The second check compares the switching-prior ablation with plain TL-VSCM at large $T/\tau$. Any claim that energy statistics should shape the causal model rests on this comparison.

The third check is downstream. The discovered graph is used to throttle predictively the block identified as causing a memory-tier hotspot, using look-ahead in the spirit of Zibaeenejad and Chen. This is compared with reactive threshold throttling and with throttling guided by correlation alone. The comparison is scored by time spent above 75°C in memory tiers and by throughput lost. A correct graph should cut overshoot relative to the correlation-guided policy without throttling bystander blocks.