Heat is a terrible communication medium, and that is exactly why we need to know its real limits. On a multicore processor, a temperature change deliberately produced by one core arrives at a neighbouring core millimetres away severely weakened and smeared out in time, because silicon stores heat and leaks it slowly — the substrate acts as a physical low-pass filter. Measured on real hardware, this attenuation is brutal: practical transmission drops from more than 45 bits per second within a single core to about 5 bits per second one core over, with theoretical capacity falling from roughly 300 to roughly 50 bits per second On the capacity of thermal covert channels in multicores, and a simpler threshold-based scheme manages only 0.33 bits per second across that same one-hop gap Thermal Covert Channels on Multi-core Platforms. These channels matter not as a replacement for wires but as a security problem: heat crosses air gaps BitWhisper: Covert Signaling Channel between Air-Gapped Computers using Thermal Manipulations and hardware isolation boundaries such as ARM TrustZone Through Fabric: A Cross-world Thermal Covert Channel on TEE-enhanced FPGA-MPSoC Systems where every conventional channel is blocked, and even a few bits per second suffices to exfiltrate a key.
Every physical validation in the literature stops at a single unassisted hop. Yet communication theory for the mathematically identical diffusion channels of molecular communication has shown that active intermediate relays — nodes that sense an attenuated signal and re-emit it — reset the distance decay and restore capacity Relaying in Diffusion-Based Molecular Communication, and theoretical nanoscale modelling shows thermal signals collapsing from 8.58 Mbps at 100 nm to 0.37 kbps at 2.5 µm precisely because no one relays them A Theoretical Modeling and Analysis of Communication via Heat Flow at Nanoscale. The gap is concrete: no one has ever built and measured a regenerative thermal relay on real silicon. This proposal does exactly that, and uses the measurement to test whether our analytical models of thermal channels survive contact with a multi-hop, relayed, real-system deployment.
Core Claim
The claim is that a chain of intermediate cores acting as decode-and-forward repeaters — each sensing the weakened thermal wave on its own digital thermal sensor, deciding the bit, and re-modulating local heat generation to pass it onward — can sustain a working link across three or more core-to-core hops at bit rates and error rates comparable to today's one-hop links, where the direct unassisted channel is effectively dead. The intuition is simple: per-hop attenuation compounds multiplicatively, so a direct H-hop link inherits the decay of every hop at once, while a relayed link only needs each single hop to work. If $a$ denotes the empirical per-hop capacity ratio (roughly $50/300 \approx 0.17$ from On the capacity of thermal covert channels in multicores), then the direct-link capacity across $H$ hops should scale like $C_{\text{direct}}(H) \approx C_1\, a^{H-1}$, while a decode-and-forward chain is limited only by its worst hop, $C_{\text{relay}}(H) \approx \min_h C_h \approx C_1$. For $H = 3$ this is a predicted gap of roughly two orders of magnitude, and the experiments below are designed to confirm or refute it on physical hardware.
The secondary claim concerns validation itself. Analytical thermal-channel models assume a linear system with stationary Gaussian noise and carefully controlled fans and frequency governors On the capacity of thermal covert channels in multicores; real machines running real workloads violate all of this, as the dynamic-thermal-management literature shows directly on silicon Reinforcement Learning-Based Inter- and Intra-Application Thermal Optimization for Lifetime Improvement of Multicore Systems, ReLeTA: Reinforcement Learning for Thermal-Aware Task Allocation on Multicore. A relay chain makes this worse, because each relay must decode a weak signal under whatever noise its own core is experiencing. We therefore build the relay decoders to be adaptive — conditioning on observable runtime state such as core frequency, fan speed, and the relay's own utilisation — and measure how much of the analytically predicted capacity survives. This turns the relay chain into a stress test of the analytical models themselves.
Research Program
Three things get built. First, a relay transceiver stack: a heat transmitter built from controllable stress workloads following Thermal Covert Channels on Multi-core Platforms; Manchester encoding with 50% duty cycle to hold the baseline temperature steady against thermal drift, as shown necessary in On the capacity of thermal covert channels in multicores; a per-relay decoder that reads the local digital thermal sensor via the Linux coretemp module and re-modulates the relay core's own workload to forward the decided bit. Second, an adaptive decoding layer: the relay's decision threshold and timing are conditioned on runtime features (current core frequency, package fan state, background load), replacing the static thresholding that is known to break under dynamic voltage and frequency scaling. Third, a measurement and validation harness that logs every hop's raw sensor traces so the analytical per-hop transfer functions can be fitted and compared against the heat-equation models the field uses, in the spirit of matching finite-element predictions to infrared thermography within tenths of a degree Quantifying heat transfer in DMD-based optoelectronic tweezers with infrared thermography] — except here for transient, time-modulated signals rather than steady-state profiles, which that literature explicitly leaves unverified.
The analysis plan treats the chain as a per-hop channel cascade: measure each hop's empirical transfer function and noise, re-estimate per-hop capacity with the water-filling method of On the capacity of thermal covert channels in multicores, and check whether the end-to-end achieved rate matches the min-over-hops prediction. Deviations localize exactly where the linear, stationary-noise assumption fails, which is itself a publishable result about the limits of analytical thermal-channel modelling.
Proposed Experiments
The primary testbed is a dual-socket Intel Xeon E5-2690 server running Linux, chosen deliberately because it is the platform on which the strongest prior result was reported: 27 bits per second achieved at under 1% error rate for same-core transmission, with practical one-hop rates above 5 bits per second and theoretical capacities near 300 and 50 bits per second for same-core and one-hop respectively On the capacity of thermal covert channels in multicores. Reusing this platform makes the comparison to the strongest baseline direct and honest. A secondary testbed, an Odroid-XU3 board with its Samsung Exynos SoC, tests portability; its sensors are near-noiseless but quantised to 1 °C steps On the capacity of thermal covert channels in multicores, which stresses the relay decoders differently.
The core experiment is a three-hop chain (cores laid out linearly on one socket, transmitter at one end, receiver at the other, two relays between). For each configuration we report achieved bit rate at a fixed 1% bit error rate target — the operating point used by the strongest baseline — plus per-hop capacity estimates and the attenuation profile from one to four hops, extending the 1-to-4-hop attenuation measurements of Thermal Covert Channels on Multi-core Platforms] from passive characterisation into active communication. A second experiment runs the chain under adversarial-but-realistic conditions: background PARSEC workloads (the suite used on physical Intel silicon in ReLeTA: Reinforcement Learning for Thermal-Aware Task Allocation on Multicore]) and enabled dynamic voltage and frequency scaling, measuring how much rate the adaptive decoders recover over static ones. A third experiment validates the model side: fitted per-hop transfer functions are compared against a finite-element heat-conduction model of the die, and the residual mismatch is quantified, following the validation methodology that caught Joule heating dominating optical absorption in Quantifying heat transfer in DMD-based optoelectronic tweezers with infrared thermography but applied to transients. Success for the headline claim is a relayed three-hop link sustaining at least 5 bits per second at under 1% bit error rate — matching today's one-hop achieved rate — while the direct three-hop link is measured at whatever near-zero rate it actually delivers; no reported result for direct multi-hop transmission exists on Lacuna, and establishing that number is itself a contribution.
Baselines
The naive reference point is direct on-off keying with static threshold decoding: the transmitter runs stress loops, the distant receiver thresholds a single sensor, with no relaying and no channel-aware processing. This is the scheme of Thermal Covert Channels on Multi-core Platforms, which achieved 0.33 bits per second across one hop on an Intel Xeon server, and it defines the floor: whatever the direct three-hop link delivers almost certainly falls below this. The strongest existing competitor is the water-filling capacity analysis with Manchester encoding and a kernel-smoothed Naive Bayes detector of On the capacity of thermal covert channels in multicores, which achieved 27 bits per second at under 1% error on the same Xeon E5-2690 platform we use, and above 5 bits per second across one hop; it is the strongest because it is the best physically validated thermal transceiver in the literature, and beating it means our relay chain must match its one-hop performance at three hops on the same machine, same sensors, same encoding, and the same tuning budget for detector hyperparameters. Both baselines and our system run on identical hardware, identical core layouts, identical Manchester encoding, and each is given the same detector-tuning sweep, so any rate difference is attributable to relaying and adaptivity rather than unequal effort. Two ablations isolate the proposal's ingredients: a passive-relay chain, in which intermediate cores detect nothing and simply heat in a fixed pattern, shows whether mere intermediate heat injection (a crude amplify-and-forward) helps or just adds interference; and a relay chain with adaptive decoding switched off (static thresholds) isolates how much of the multi-hop gain comes from channel-aware runtime conditioning versus from regeneration alone. Beating the strongest baseline means sustaining at least 5 bits per second at 1% error across three hops; beating the ablations means the full system's rate exceeds both the passive-relay and static-threshold variants by a margin large enough that per-hop error propagation, not detector tuning, explains the difference.
Why not just…?
Why not just slow the symbol rate instead of building relays? Stretching each bit longer does reduce the overlap between consecutive symbols, and the nanoscale analytical model leans on exactly this assumption A Theoretical Modeling and Analysis of Communication via Heat Flow at Nanoscale]. But slower on-off signalling makes the drift problem worse, not better: without Manchester encoding's forced 50% duty cycle, baseline chip temperature wanders and static threshold decoders fail outright, as shown empirically in On the capacity of thermal covert channels in multicores]. Slowing down buys error rate, never reach: it cannot recover the exponential-in-hops amplitude loss, and our direct-link attenuation measurements from one to four hops will quantify exactly how little it buys. That said, we will report the direct link at its own optimally slow symbol rate, so the comparison is against the best simple signalling can do, not a strawman.
Why not just fuse many sensors (SIMO) instead of relaying? This is the most serious alternative, and it is cheaper: no relay protocol, no per-hop decoding delay, no error propagation across hops. BitWhisper: Covert Signaling Channel between Air-Gapped Computers using Thermal Manipulations] explicitly hypothesised multi-sensor fusion as the way to fight attenuation, and multi-type diversity provably boosts capacity in diffusion channels Relaying in Diffusion-Based Molecular Communication]. Our honest position is that combining across the receiver's nearby sensors can reject noise and modestly extend reach, but it cannot reset distance: sensors beyond the first hop see the same compounded attenuation the receiver does, and no linear combination of weak, ISI-smeared copies regenerates a symbol that never arrived distinguishably. We will settle this empirically rather than argue it: a maximal-ratio-combining receiver over all sensors on the socket is included as an additional baseline, and the decisive check is whether sensor fusion alone sustains a three-hop link at meaningful rate. If it does, the relay's value shrinks to hop counts beyond three, and we will report that plainly.
Why not just simulate with HotSpot or COMSOL instead of building physical rigs? The validation literature gives direct evidence that simulation alone misleads in exactly this setting: finite-element predictions matched steady-state thermography to within tenths of a degree in Quantifying heat transfer in DMD-based optoelectronic tweezers with infrared thermography], yet physical measurement was still needed to discover that the assumed heat source was wrong; the authors of Reinforcement Learning-Based Inter- and Intra-Application Thermal Optimization for Lifetime Improvement of Multicore Systems] rejected offline simulators outright as too slow and blind to OS-level runtime interference; and physical experiments revealed wall-thickness degradation that simulation missed in Mechanism of Mechanical Analysis on Torsional Buckling of U-Shaped Bellows in FLNG Cryogenic Hoses]. Our relay chain is precisely the case simulation handles worst: a cascade of quantised sensors, per-relay decoder errors, and non-stationary DVFS noise. We do use a finite-element model — but as the object being validated, running side by side with the physical chain, so the paper reports where the model diverges from silicon rather than substituting for it.
Decisive Checks
The proposal stands or falls on three measurements. First, the three-hop relayed link versus the direct three-hop link on the Xeon platform at matched error rates: the min-over-hops capacity prediction says the gap should be roughly two orders of magnitude, and finding anything materially smaller would falsify the core claim and would itself recalibrate how the field models multi-hop thermal attenuation. Second, the relay-versus-SIMO comparison: if multi-sensor fusion alone matches the relay chain, the relay's contribution is not worth its complexity and the honest conclusion is a fusion-based design. Third, the adaptive-versus-static decoder comparison under PARSEC background load and enabled frequency scaling: if runtime-conditioned decoding does not measurably recover rate, then the analytical models' stationarity assumptions are more robust on real servers than the dynamic-thermal-management evidence suggests, and that boundary should be mapped rather than assumed. Every raw sensor trace and fitted transfer function will be released, since the field currently has no shared physical traces for multi-hop thermal channels — all prior datasets were generated privately per paper, and no public benchmark dataset exists on Lacuna for this problem.
References
- On the capacity of thermal covert channels in multicores
- Thermal Covert Channels on Multi-core Platforms
- BitWhisper: Covert Signaling Channel between Air-Gapped Computers Using Thermal Manipulations
- Through Fabric: A Cross-world Thermal Covert Channel on TEE-enhanced FPGA-MPSoC Systems
- Relaying in diffusion-based molecular communication
- A Theoretical Modeling and Analysis of Communication via Heat Flow at Nanoscale
- Reinforcement Learning-Based Inter- and Intra-Application Thermal Optimization for Lifetime Improvement of Multicore Systems
- ReLeTA: Reinforcement Learning for Thermal-Aware Task Allocation on Multicore
- Quantifying heat transfer in DMD-based optoelectronic tweezers with infrared thermography
- Mechanism of Mechanical Analysis on Torsional Buckling of U-Shaped Bellows in FLNG Cryogenic Hoses