# IGZO Edge Systems: Benchmarking, Energy and Safety Technical Report

Version 3.1 · 12 October 2026 · IGZO Evidence Atlas

## Purpose

This document specifies how analogue in-memory computing should be tested in an edge task. It claims no measured system advantage for the project. Device, circuit, algorithm, and complete-system boundaries are kept separate. The MLCommons Tiny framework provides a methodological example for matched accuracy, latency, and energy comparisons; no MLPerf result is attributed to this project [1].

## 1. Ideal model and physical array

For an ideal crossbar, column current is I_j = Σ_i V_i G_ij. In a physical array, electrode and line resistances, sneak paths, state variability, read noise, drift, and ADC quantisation cause errors. Xia and Yang [2] and Jiang and colleagues [3] discuss this layered error budget. The ideal model is a starting control, not a measured array result.

Passive crossbars and 1T1R should be assessed with the same device material, matched accuracy target, and comparable area and peripheral boundaries. The selector transistor's area and control energy must be included. Weight encoding, write verification, and recalibration frequency are recorded for the matched task.

## 2. Energy and latency boundaries

Programming: E_write = Σ_pulse ∫ V(t)I(t)dt + E_verify + E_retry.

Inference: E_infer = [∫_(t0)^(t1) P_system(t)dt] / N_valid. N_valid is the number of inferences meeting the preregistered validity rule. Sensor and preprocessing, DAC/drivers, array, ADC/readout, controller, memory, and actuator must also be reported as distinct boundaries. Latency runs from the sensor sample timestamp to an available decision; report p50, p95, p99 and the count of worst-case events.

Array energy alone is not system energy. ADC/DAC, calibration, and idle power may dominate. PUE is a facility-level measure and cannot be calculated from cell energy. IEA electricity data supplies global context, not proof of savings in this project [4].

## 3. Fair baseline

Compare devices on the same task, data split, input resolution, accuracy target, throughput, latency requirement, and thermal environment. If one side uses calibration during training, define a comparable processing boundary for the other. Report parameter count, numerical precision, and confidence interval. GPU, FPGA, MCU, or CPU can be a baseline, provided the complete power interval and duty cycle are declared. A figure such as “300×” is not a result without a matched experiment.

## 4. Four demonstrations and safety gates

1. **Motorsport telemetry:** Start with recorded data and session/vehicle-separated training and test sets, then shadow mode. Report event recall, false alarms, and p99 latency.
2. **Smart grid:** Use synthetic/open events for decision support. A protection relay function requires independent safety evaluation.
3. **UAV:** Separate simulation, bench, tethered, and controlled flight stages. Define safe behaviour for GPS loss and sensor degradation.
4. **Artificial nociceptor:** Test stimulus-threshold-response curves and hysteresis. This is not biological or clinical equivalence; printed IGZO nociceptor research supplies only relevant technical context [5].

## 5. Security and risk management

Inject faults for unauthorised weight writes, version rollback, power loss, corrupted calibration, and sensor attack. Define expected safe state, audit trail, and recovery time in advance. NIST AI RMF is a voluntary framework for mapping, measuring, and managing risks; it is not product certification [6]. Patent-specific restricted writing cannot replace authorisation or attack-resilience testing.

## 6. Reproduction package

Preserve data-set version, training/validation/test split, device and batch counts, raw sensor and electrical traces, timestamp definition, calibration version, software commit, baseline configuration, failed runs, and uncertainty method. For confidential raw data, disclose the sharing constraint and provide a reproducible synthetic proxy where possible.

## References

1. MLCommons, MLPerf Tiny, https://mlcommons.org/2021/06/mlperf-tiny-inference-benchmark/
2. Xia and Yang, *Nature Materials* (2019), https://www.nature.com/articles/s41563-019-0291-x
3. Jiang et al., *Nature Materials* (2026), https://www.nature.com/articles/s41563-026-02600-y
4. IEA, *Energy and AI* (2025), https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
5. *Fully printed IGZO memristor arrays for artificial nociceptors*, *Science China Materials* (2024), https://doi.org/10.1007/s40843-024-2887-9
6. NIST AI RMF 1.0, https://doi.org/10.6028/NIST.AI.100-1

## Portfolios 1–2 facility boundary (v3.1)

The portfolios' 300× processor efficiency, 90% cooling reduction and PUE≈1.0 are not verified integrated IGZO-system results. Separate chip, server and facility electricity measurements require a matched workload, accuracy target and time window. The IEA estimates about 415 TWh of global datacentre electricity in 2024; the cooling share ranges from about 7% in efficient hyperscale centres to over 30% in less-efficient enterprise facilities. A universal 40–50% share is inappropriate. Source: https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai