Short Learned Precoders for Over-the-Air Majority Votes Inside Reverberant Chip Packages

Large AI accelerators are now built from many small dies, called chiplets, placed in one package. In these systems, moving data between chiplets rather than doing arithmetic often sets the run time. Collective operations dominate that traffic: one example is summing or majority-voting the same vector across many chiplets and sending the result to many destinations. One proposed shortcut is over-the-air computation: every chiplet transmits at the same instant on a shared on-package radio channel, and the radio waves add up physically, so a receiver reads the sum or majority vote in one time slot instead of K. The catch is that the inside of a metal chip package behaves like an echo chamber. Each transmitter's pulse reaches the receiver as a long smear of reflections. The waves then add up to garbage unless every transmitter pre-shapes its signal so that all the smears line up at the receiver at the same moment.

This note proposes to learn those pre-shaping filters, called precoders, jointly for all K transmitters. The filters are deliberately short so that they can run at realistic on-chip sample rates. The training objective has two extra terms. The first penalizes energy that spills onto other receivers and adjacent frequency bands (spectral leakage) that concurrent links are using. The second averages performance over an ensemble of perturbed channels and transceiver impairments rather than a single nominal simulation. The core claim is a design rule. Precoders tuned on one clean electromagnetic simulation, or tuned only for the aggregation receiver, will look best in simulation and fail first when the package drifts or the neighbors are switched on. A short, jointly trained, leakage-aware and impairment-robust precoder will keep the over-the-air vote accurate at lower transmit power and higher symbol rate than the existing alternatives.

What exists and what is missing

Three lines of work bracket this idea, and none of them closes it. WHYPE is the work that motivates the proposal. It maps the "bundling" step of hyperdimensional computing, a bitwise majority vote across many long binary vectors, onto an in-package wireless broadcast where phase-shifted BPSK signals superimpose. It reports a 3.2x interconnect-area reduction against a wired reduction tree. Its channel, however, is static, flat and non-dispersive, its front ends are ideal, and it notes that error rates depend on where the receiver sits without addressing that.

The time-reversal papers measure the channel WHYPE ignores. Collective Communication Patterns Using Time-Reversal Terahertz Links at the Chip Scale, Time Reversal for Near-Field Communications on Multi-chip Wireless Networks and Towards Scalable Multi-Chip Wireless Networks with Near-Field Time Reversal use full-wave simulation of a 140 GHz package with 12 antennas. They show that pre-filtering with the time-reversed channel response turns reverberation into a focused peak, giving 8 dB higher interference-limited signal quality and about 50 Gb/s per link. These papers supply the method we start from. They also constrain the design in two ways. First, they use time reversal to keep transmissions apart, not to make them add. Second, they flag that full time-reversal filters need sampling at tens of GHz, which is a serious hardware burden.

The third line is the optical transceiver autoencoders of Model-Based End-to-End Learning for WDM Systems With Transceiver Hardware Impairments and End-to-end Autoencoder for Superchannel Transceivers with Hardware Impairment. These jointly learn the mapper, the pulse-shaping filter, the predistortion and the demapper under modeled modulator, DAC and amplifier impairments. They report 37–50% smaller guard bands and up to 60% lower symbol error than root-raised-cosine filtering. They also found a "single-channel trap": a transceiver trained alone learns to leak into its neighbors' bands. We reuse their modular, model-based training. We expect an analogous trap here, where a precoder trained only to serve the aggregation receiver spills energy onto other receivers.

Macro-scale over-the-air computation supplies the function and its theory. One-Bit Over-the-Air Aggregation runs a one-bit majority vote over cyclic-prefix OFDM with truncated channel inversion, and shows that error from imperfect channel knowledge decays as $1/\sqrt{K}$ in the number of devices. It assumes macro-cell channels and ideal radios, and it uses rectangular-pulse OFDM, which has high spectral sidelobes, as Cyclic block filtered multitone modulation documents.

The obvious combination is "time reversal plus over-the-air majority in a package." Even that has not been evaluated in the material found. This proposal differs from it in three specific ways: the precoders are designed jointly across transmitters rather than independently, they are length-limited to a hardware budget, and they are trained against leakage and channel drift.

Core Claim

Let $s_k[n]\in\{-1,+1\}$ be the bit chiplet $k$ wants to vote, $g_k$ its precoder (an FIR filter of $L$ taps), $h_{k,r}$ the channel from transmitter $k$ to receiver $r$, and $w$ additive noise. The aggregation receiver $r^\star$ sees $y_{r^\star}=\sum_k h_{k,r^\star} * g_k * s_k + w$ and outputs the sign. We train all $g_k$, plus optional lightweight predistortion, to minimize

$$ \mathbb{E}_{\theta\sim\Theta}\Big[\Pr\big(\mathrm{sign}(y_{r^\star}) \neq \mathrm{maj}(s)\big) + \lambda \sum_{r\neq r^\star} P_{\text{leak},r}\Big], $$

Here $\theta$ indexes a perturbed package-and-hardware instance drawn from an ensemble $\Theta$: geometry and permittivity variations, per-chiplet oscillator phase and frequency offsets, amplifier compression, and DAC resolution. $P_{\text{leak},r}$ is the power delivered to receiver $r$ in the time window and band used by a concurrent link, and $\lambda$ trades vote accuracy against harm to neighbors.

The claim is that, at a tap budget $L$ far below full time-reversal length, this precoder beats every alternative listed under Baselines on four measures at once: vote accuracy on held-out perturbed packages, leakage, and energy per reduced bit at a target end-task accuracy.

Hyperdimensional computing tolerates high bit-error rates; WHYPE reports Omniglot few-shot accuracy above 99% even at a bit-error rate of 0.26. The payoff of robustness is therefore not raw accuracy. It is how much transmit power or symbol rate can be traded away before the end task degrades. We make that the headline metric.

Research Program

The first stage builds the channel ensemble. We reproduce the 11×11 mm², 12-antenna, 140 GHz package geometry of the time-reversal papers in CST Microwave Studio and extract impulse responses for all transmitter–receiver pairs. We then generate perturbed instances by shifting antenna positions, changing dielectric constants within a thermal range, and moving or adding a heat-spreader lid. One subset of perturbations is held out entirely for testing.

The hardware impairment models come from the cited circuit literature. Amplifier power and compression across frequency follow Frequency-Dependent Power Consumption Modeling of CMOS Transmitters for WNoC Architectures. DAC resolution follows the 6-bit effective-resolution setting of the optical autoencoder work. Oscillator phase noise and frequency offset are modeled independently per chiplet, since chiplets do not share an oscillator.

The second stage trains precoders by differentiating through these models, following the model-based autoencoder recipe. Where a component is not differentiable, we fall back to the policy-gradient variant used in the same optical work. We sweep the tap count $L$, the number of voters $K$, and the leakage weight $\lambda$.

The third stage connects to workloads. We run Omniglot bundling through a WHYPE-style simulated pipeline. We also run one-bit majority aggregation of gradient signs for data-parallel training across chiplets on MNIST and CIFAR-10, mirroring the One-Bit Over-the-Air Aggregation setup. Communication time is then translated into end-to-end time using the inter-chiplet traffic shares that the GEMINI-based characterization reports for 18-chiplet packages, where package-level communication is 72.6% of Transformer execution time.

The sim-to-silicon gap enters as an evaluation discipline rather than a fabricated chip. We claim no silicon. Every result is reported on the nominal channel and on held-out perturbed instances, and the gap between the two numbers is a primary outcome.

Baselines

The naive reference point is a wired reduction tree on the package network, the baseline WHYPE compares against. We cost it with WHYPE's area accounting and with the GEMINI multicast fan-out statistics, which show over 30% of LSTM/GNMT multicasts reaching 12 destinations. A wireless scheme that cannot beat this tree on latency per reduction at equal end-task accuracy has no reason to exist.

The second reference is WHYPE's own scheme run on the realistic channel: plain BPSK with one complex gain per transmitter for phase alignment. This tests whether HDC's error tolerance makes channel shaping unnecessary. No reported result was found on Lacuna for WHYPE on a dispersive package channel.

The strongest existing method is independent full-length time reversal. Each transmitter pre-filters with its own time-reversed response toward the aggregation receiver, so that all focal peaks coincide. This is the natural extension of the time-reversal papers, which report an 8 dB SINR gain and near error-free 50 Gb/s for unicast. No reported result was found on Lacuna for time reversal used for over-the-air aggregation. We also run it truncated to the same $L$ taps as our precoder.

A further baseline is OFDM-based one-bit aggregation transplanted into the package, with the cyclic prefix sized to the measured delay spread and truncated channel inversion. Its paper reports $1/\sqrt{K}$ bias decay but no in-package figure. We run it with the same amplifier and DAC models, so that its peak-to-average power is penalized fairly. All methods share the channel ensemble, the impairment models, the transmit power grid, and a matched hyperparameter search budget.

Beating the strongest baseline means two things. At equal $L$ and equal held-out perturbation, our precoder reaches the end-task accuracy target with at least 3 dB less transmit power, or at a higher symbol rate. It must also deliver lower leakage to concurrent receivers.

Three ablations isolate the ingredients: removing the leakage term, training on the nominal channel only, and replacing joint training with independently trained per-transmitter filters.

Why not just…?

Why not just keep the reduction wired, or put adders in the network switches? In-network reduction is effective at cluster scale: In-Network Collective Operations reports an 8 GiB AllReduce on 1,024 nodes dropping from 352 ms to 151 ms. The same paper flags that wide accumulators are needed for low-precision sums and that the gains shrink by Amdahl's law. A one-bit majority sidesteps the accumulator growth. Whether the wireless version wins in a package is an empirical question that our latency-per-reduction comparison against the wired tree settles. If the wired tree wins at 18 chiplets, we report that.

Why not just send WHYPE's plain BPSK, since hyperdimensional computing already tolerates a 0.26 bit-error rate? This is the most serious threat. The 0.26 figure comes from random, independent bit flips. A reverberant channel instead produces errors that depend on receiver position and on neighboring bits, and WHYPE itself observed the position dependence. The single-gain baseline above measures this directly. If it matches our precoder on held-out packages at equal power, the channel-shaping part of this proposal is unnecessary for HDC. The work would then narrow to the gradient-sign workload, which is less error-tolerant.

Why not just use independent time reversal, which already focuses energy? Full-length time reversal requires sampling at tens of GHz, as its authors note. Independent focusing also ignores the cross-terms among the K transmitters and the energy left at other receivers, including the symmetric "hotspots" those papers report. The truncated-time-reversal and independent-training comparisons show whether joint, short training matters. If full time reversal truncated to $L$ taps does as well, the contribution reduces to the leakage and robustness terms.

Why not just use OFDM with a cyclic prefix, as macro-scale over-the-air aggregation does? In a long-delay-spread cavity, the cyclic prefix becomes a large overhead. Rectangular pulses leak into neighboring bands, which Cyclic block filtered multitone modulation shows costs up to 10 dB at a symbol error rate of $10^{-4}$ in frequency-selective channels. OFDM's high peak-to-average power also collides with the nonlinear amplifiers of the circuit models, and the FFT blocks add area. The transplanted-OFDM baseline is run under these same costs. If it wins on energy per reduced bit, the right follow-up is a pulse-shaped multicarrier design rather than time-domain precoding.

Decisive Checks

Three checks decide the proposal. The first is the robustness gap: the precoder trained on the nominal channel only must lose noticeably more accuracy on held-out perturbed packages than the ensemble-trained one. If it does not, the sim-to-silicon concern is overstated for this setting.

The second is a reduction-level trap. We expect, though we have not shown, an analogue of the single-channel trap. Removing the leakage term should increase power delivered to concurrent receivers while barely improving vote accuracy, which would confirm that neighbor-aware training is necessary.

The third is energy at equal task quality. Using the frequency-dependent transmitter power model, the proposed precoder must reach the Omniglot and CIFAR-10 accuracy targets at lower energy per reduced bit than both truncated time reversal and transplanted OFDM, across $K$ from 3 to 16. A further sub-check is whether vote error falls with $K$ at the $1/\sqrt{K}$ rate predicted for channel-estimation bias in the one-bit aggregation analysis.

If all three checks hold, the result is a concrete, hardware-budgeted recipe for in-package over-the-air reductions. If the HDC-tolerance check fails, the result is still a clean negative finding: a statement of when channel shaping is unnecessary.