The Unsonic Project

01 — the productUnsonic on the air

Silence is the most expensive thing a broadcast chain can transmit. An unmodulated channel is not neutral: it opens the noise floor, invites hiss and hum into the listener's speakers, and trips the silence-detection alarms that sit downstream in every automated plant. Unsonic is the audio bed that occupies that space — engineered content that keeps the channel carrying signal when the programming has nothing to say.

It runs on station / network name, market, where it has been in rotation since first air date. It is ordinary audio: it goes down the same path as everything else on the chain, through the console and the processor, and it is heard — or rather, arranged so that it is not. Nothing about the on-air product touches RF. The catalog is registered with ASCAP under Xtra Ad Service, Inc., and the work is published through the Unsonic Audio channel.

That is the part of this business that already works, and it is what pays for the rest.

02 — the benchWhirr, and the question it asks

Unsonic began as a signal. Whirr asks whether that signal can do physical work.

The premise: an ESP32-S3 drives an array of GPIO pins — rods — from its I²S peripheral in PDM mode, at rates far above hearing. Each rod is a pin switching a small capacitance millions of times a second. The question was whether an array of them, arranged and driven deliberately, changes how heat leaves the die. If it did, the implication would be substantial: a thermal-management effect with no moving parts, no heatsink, and no bill of materials beyond pins the designer already has.

To answer that, the rod count had to be the only thing that changed. Everything else — load, ambient, settle state, firmware — had to be pinned down and proven pinned down. Most of the work went there.

The instrument

The ESP32-S3's internal temperature sensor quantizes to 1 °C. That is roughly two hundred times coarser than the effect worth looking for, so the sensor had to be made better before the experiment could mean anything.

• Dither correction. The quantizer's own noise, averaged correctly across a long run, recovers an effective resolution of 0.002–0.005 °C from a sensor that reports whole degrees.
• Lock-in detection. The array is modulated on and off at 1/600 Hz and the temperature record is demodulated at that frequency, which rejects room drift by construction rather than by hope.
• Serial-commanded rod count. Firmware (Node 2.4) takes rod count, load state, and verification handshake over the wire, so a configuration change no longer means a reflash — and the host records which configuration actually ran rather than which one was intended.
• Step-response fitting. A level-crossing estimator recovers time constants from heavily quantized data where a naïve exponential fit does not survive.

This rig is the durable asset of the project. It is a general instrument for measuring milliwatt-scale thermal effects on a commodity part using only that part's own sensor.

03 — the findingWhat fifteen runs said

The primary metric was tau_cool, the cooling time constant from the recovery phase, chosen because it is independent of ambient temperature. Fifteen runs across three sessions, comparing a five-rod array against a ten-rod array.

The run that was almost a result

The 31 July session showed a clean ten-rod advantage. It was an artifact. Every five-rod run had gone in the morning and every ten-rod run in the afternoon, and room temperature tracked time of day with R² = 0.94 — the schedule, not the array, was doing the work. Interleaving the configurations on 4 August erased the effect entirely.

That catch is presented here deliberately. The same session log also documents hot-start bias, mid-recovery ambient drift, HVAC cycling after plateau, under-saturated load windows, and a firmware default that started runs already under load. Each was found by the instrument and corrected in the protocol.

Two things the null result exposed

Chasing the absent cooling effect turned up two facts about the part that are more useful than the hypothesis was.

The array is a heat source, and a measurable one. Lock-in detection found 0.9–1.32 °C at SNR 8.5–11.0 between rods-off and rods-on — about 25 mW. A capacitive-switching model of the pins predicts 1–7 mW. The excess is the I²S peripheral and its DMA engine, and it is very nearly independent of how many pins the array drives.

The thermal system is two-pole, not one. Step response resolves a fast constant near 101 s and a slow one near 283 s. The single-pole tau_cool ≈ 123 s used throughout was serviceable but never exact, which explains residuals seen from the first session onward.

04 — the positionWhat this is and is not

Whirr set out to test a cooling claim. The bench does not support one, and this page will not make one. Driving the rod array raises die temperature rather than lowering it, and adding rods past five does not change the outcome within the resolution of a rig that resolves thousandths of a degree.

What the work produced instead:

The honest framing for a semiconductor reviewer is this: the hypothesis under test was firmware and board-level, on a part somebody else fabricated. It did not survive contact with the instrument. The instrument survived, and it is the piece that could become silicon — an on-die thermal characterization block with dither correction and lock-in demodulation in hardware, aimed at the calibration and binning problem that every SoC with a coarse internal sensor already has.

05 — the askWhere a screening panel comes in

Silicon Catalyst evaluates applicants against eleven criteria and puts each one in front of a wide group of ecosystem reviewers before admission to the 24-month program. This page is submitted as evidence toward the technical and team criteria specifically, and it is deliberately a negative result presented in full.

The claim being made is not that Whirr works. It is that this bench asks a question properly, instruments it beyond the resolution the vendor provides, catches its own confounds before publishing them, and reports against its own interest when the data says so. That disposition is the thing being offered — funded to date by share of Unsonic revenue directed to the bench against bench cost to date, with no outside capital.

The next test is specified and waiting: isolate per-pin dissipation by running the lock-in against a five-pin floor to cancel the I²S/DMA term, and fit the two-pole model directly rather than inheriting a single-pole approximation. Full run logs, firmware, analysis toolchain, and the benchmark report are available on request.