The Energy Case for Tier-0: A Pre-Registered, Measured Energy Frontier and System-Gating Law for a Frozen Decision Token on Microcontroller Silicon
Year: 2026
Type: preprint
Abstract: The Energy Case for Tier-0: A Pre-Registered, Measured Energy Frontier and System-Gating Law for a Frozen Decision Token on Microcontroller Silicon Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: 10.5281/zenodo.24322465 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign · a characterization of previously published methods that are the subject of filed U.S. patent applications, measured on commodity microcontroller silicon against third-party prior-art baselines; it discloses no new algorithmic subject matter and is accompanied by no new patent application — see the notice below. Abstract The prevailing efficiency argument for a cheap edge classifier is made in FLOPs. FLOPs are an estimate; the quantity a battery pays is energy, and on a cache-poor microcontroller energy is set as much by memory traffic and duty cycle as by operation count. This paper measures, rather than estimates, the energy of a frozen, deterministic, class-discriminant single-token decision encoder against the efficient-edge frontier, on real Nordic silicon (a Cortex-M33, nRF54L15, instrumented with a Power Profiler Kit II), under two pre-registered campaigns with frozen Go/No-Go bands. The first establishes a measured energy-per-decision frontier: the 1-byte token decides in 12.3 µJ, against 19× (depthwise-separable CNN), 29× (a MicroNets-style micro-CNN), 135× (ROCKET), and 481× (an int8 "SRAM-wall" CNN the 32-bit path cannot even fit) more energy for the raw-signal neural and classical time-series frontier — a span of roughly two and a half orders of magnitude, with the token competitive or better on accuracy. A host study across ten datasets confirms the compute-proxy version (the token is Pareto-non-dominated by the neural frontier at 20–300× fewer multiply-accumulates at matched accuracy, 10 of 10) and, reported straight, its honest boundary: a trivial linear model on the same cheap features ties the token on accuracy-per-operation, so the token's defensible value over a linear baseline is not accuracy-per-FLOP but the invariant package — a deterministic ≤256-entry lookup, a one-byte output, and a built-in label-free nearest-centroid competence monitor — and the compute/energy win is specifically against the neural frontier the field reaches for. The second campaign composes the measured per-tier energies into a system-gating law: a frozen Tier-0 token in front of an expensive Tier-N stage is an energy valve that saves total system energy only below a break-even wake-rate, a closed form r* = 1 − E₀/E_N in the measured per-tier energies. We confirm the composition on silicon: an interleaved Tier-0+Tier-N sweep shows the composed per-decision energy is linear in the wake-rate to better than 0.1 % (R² = 1.000000), so the composed system costs exactly E₀ + w·E_N with no context-switch or cache overhead, and the measured break-even matches the closed form to within 0.1 % for two on-device Tier-N magnitudes. Translating the measured energies into the battery-swap/maintenance interval that dominates deployed-sensor ownership, the token-gated architecture runs for months on a coin cell where always-on deep learning runs for hours, and an operating-mode × domain map (literature-grounded cadences for seizure EEG, arrhythmia ECG, bearing vibration, wearable activity, and continuous glucose monitoring) shows that in continuous rare-event monitoring the token-triggered architecture is the only one delivering full event coverage and multi-month battery life. We then red-team the frontier across seven axes and report the results, including the ones that temper the claim: the advantage is window-length- and channel-count-dependent (≈5–29× versus the smallest CNN, 24–481× versus a real CNN, never inverting); every model's latency is bounded with under-0.2 % jitter; the ratios are preserved on a second core (slightly larger on a Cortex-M4F); the naive int8 baseline is already at the scalar-compiler ceiling (fair, not a strawman); and — a correction folded into a companion erratum — on this hardware's DC-DC buck regulator the supply energy per decision is approximately voltage-independent (≈ −8 % at 1.8 V, not the quadratic saving a linear-regulator intuition predicts). This is a characterization, on commodity hardware, of previously published, patent-pending methods against third-party prior-art baselines; it discloses no new algorithmic subject matter and is accompanied by no new patent application. The per-deployment operating-point and architecture-selection procedure is retained as trade secret. Highlights · A measured energy-per-decision frontier, not a FLOP estimate. On one Cortex-M33 the 1-byte token decides in 12.3 µJ, versus 19–481× more for the raw-signal neural / classical time-series frontier (depthwise-separable 235 µJ · micro-CNN 352 µJ · ROCKET 1.66 mJ · int8 CNN 5.92 mJ) — all at a constant, directly measured active power — with the token competitive-or-better on accuracy. · The honest boundary, reported straight. A linear model on the same cheap features ties the token on accuracy-per-operation; the token's defensible edge over a linear baseline is the invariant package (deterministic ≤256-entry lookup, one-byte output, built-in label-free monitor), and the energy win is specifically versus the neural frontier — not a claim to be the cheapest classifier that exists. · A system-gating law, confirmed on silicon. Composing the token gate with an expensive Tier-N stage is linear in the wake-rate to <0.1 % (R² = 1.000000) — no composition overhead — so the break-even wake-rate r* = 1 − E₀/E_N predicts the measured crossover to within 0.1 % for two on-device Tier-N magnitudes. · From joules to cost: a deployability flip. At 10 Hz on a coin cell the token-gated architecture runs ≈5 months between battery swaps while always-on deep learning runs ≈8 hours; across verticals by their literature-grounded cadence, token-triggering is the only architecture with full event coverage and long battery life for continuous rare-event monitoring — and the advantage is honestly moot at ultra-low cadence (e.g. glucose), where the idle floor dominates. · A worst-case, cross-core, fair-baseline frontier. Every model's latency is bounded with < 0.2 % jitter; the token-versus-neural ratios are preserved on a second core (larger on a Cortex-M4F: 33× / 517×); and compiler optimization does not improve the naive int8 baseline (it is already the scalar ceiling), so the comparison is fair, not a strawman. · Honest scoping and a self-correction. The advantage is window-length-dependent (≈5–29× versus the smallest CNN) — 29× is the long-window best case, never an inversion; and on this DC-DC-regulated part the supply energy is approximately voltage-independent, correcting a companion deposit's linear-regulator voltage estimate. No new mechanism is introduced. The frontier, in two measured tables On one Cortex-M33 the active power is directly measured constant at 2.58 mA (7.74 mW) at 3.0 V across every model, so per-decision energy is the measured active power divided by the measured throughput and any two models' energy ratio is purely their measured throughput ratio. The measured ratio exceeds the multiply-accumulate ratio, because the networks are memory-bound over large activation buffers while the token's cost is a cache-resident feature front-end. The feature-fed linear and MLP heads tie the token, locating the honest boundary; the energy win is specifically against the neural frontier. Model Input Latency/decision Energy/decision × token LR (features) 16 feats 1.37 ms 10.6 µJ 0.9× tiny-MLP (features) 16 feats 1.39 ms 10.8 µJ 0.9× token (1 byte) 16 feats 1.58 ms 12.3 µJ 1.0× dwsep-CNN (raw) 1024 samples 30.41 ms 235.3 µJ ~19× micro-CNN (raw) 1024 samples 45.05 ms 352.0 µJ ~29× ROCKET-lite (raw) 1024 samples 214.26 ms 1658.0 µJ ~135× int8 small-CNN (raw) 1024 samples 755.96 ms 5919.0 µJ ~481× Translating the measured energies into the battery-swap/maintenance interval that dominates deployed-sensor ownership gives a deployability flip: at a 10 Hz always-on cadence on a standard coin cell, the token decides for months where always-on deep learning lasts hours. The advantage grows with cadence and ceilings below ~0.8 Hz, where the model-independent idle floor (measured 9.7 µW deep-idle) dominates and all architectures converge; the absolute lives exclude sensor and radio acquisition energy (both model-independent) — the robust claim is the ratio. Model Energy/decision Coin-cell life @10 Hz vs token token (1 byte) 12.3 µJ ~145 days (~5 months) 1.0× micro-CNN (raw) 352.0 µJ ~5.5 days ~27× shorter int8 small-CNN (raw) 5919.0 µJ ~7.8 hours ~446× shorter The system-gating law, confirmed on silicon A frozen Tier-0 token in front of an expensive Tier-N stage is an energy valve: per window it costs E_gate = E₀ + w·E_N, where w = λ·TPR + (1−λ)·FPR is the wake-rate, so gating beats always-running Tier-N exactly when w < r* = 1 − E₀/E_N. A firmware sweep that interleaves the always-on token with a 1/k-fraction Tier-N invocation shows the composed per-decision time is linear in the wake-rate to better than 0.1 % (R² = 1.000000), with intercept equal to the token's own time and slope equal to the Tier-N time — no measurable context-switch or cache penalty. Because the active power is model-independent, a linear composed time is a linear composed energy. The measured break-even matches the closed form to within 0.1 % for two on-device Tier-N magnitudes (micro-CNN: measured 0.9653 vs. closed-form 0.9651; int8 small-CNN: 0.9980 vs. 0.9979); a radio-wake Tier-N (measured 16.6 µJ BLE transmit) places the break-even at an interior wake-rate of 0.
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235491929
