The evidence chain and experimental programme

Experiment

A measured observation with sample structure, conditions and uncertainty.

Model

A simulation or estimate with explicit parameters and validation domain.

Target

A proposed pass threshold awaiting project-specific measurements.

A · Source devices

Use externally fabricated IGZO devices. Retain incoming QC, integration, calibration, and edge demonstration in the programme. Draft planning window: 18–30 months; 24 months is an example, not a guarantee.

B1 · Fabricate internally

Document substrate, film, electrode, patterning, annealing, packaging and device test in one revision-controlled chain. Requires appropriate process and safety infrastructure. Draft window: 24–36 months.

B2 · Subcontract fabrication

Specify measurable delivery conditions, lot provenance and independent incoming QC. Keep driver integration and calibration in-house. Draft window: 24–36 months.

WP0 · Protocol

Lock task, data version and split, batch and device counts, instrument settings, digital baseline, metrics and exclusions before measurement.

WP1 · Device QC

Capture stack, geometry, raw I–V, switching distributions, multilevel overlap, retention, endurance, read disturb, temperature and rejects.

WP2 · Array and circuit

Compare passive crossbar and 1T1R, including line resistance, sneak paths, selection costs, ADC/DAC, reference cells, area and raw versus calibrated MVM error.

WP3 · Algorithm

Record weight encoding, pulse and verification policy, piecewise-linear update correction, retraining and recalibration time/energy.

WP4 · Demonstrator

Report sensor-to-decision p50/p95/p99 latency, energy per valid inference, idle power, accuracy and environmental sensitivity.

WP5–6 · Security & publication

Inject unauthorised writes and faults; publish failure cases, raw traces, code revisions, uncertainty and independent replication instructions.

Datacentres and cooling

Claims of 300× processor efficiency, 90% less cooling and PUE≈1.0 are project assertions. Chip, server and facility boundaries must be measured separately. The IEA estimates 415 TWh of global datacentre electricity in 2024 and a cooling share ranging from about 7% to over 30% by facility type.

Telemetry and cognitive interfaces

Motorsport, driver-state and brain–machine interface narratives are candidate research uses. The portfolios do not provide independent end-to-end validation of 1 ms, 1 W, 99% accuracy or deployment for this IGZO platform.

Test first on recorded data with participant/session separation and shadow-mode evaluation.

Device-specific IGZO evidence

Wang et al. (2026) report a 1-kb Pt/IGZO/SiNₓ/Ta array with endurance above 10⁶ cycles and retention above 10⁶ s. That result belongs to a different device stack and does not establish the portfolio platform's performance.

Classification accuracy

Naqi et al.'s CNN and SNN percentages have a model/simulation context; they are not measured end-to-end accuracy for the proposed platform.

Retention and cycling

Retention from Martins and Naqi and the 500-cycle Naqi result belong to different devices or measurements. They cannot be merged into a universal IGZO specification.

Latency, power and 300×

Supplied drafts provide no raw sensor-to-actuator timestamps, complete power trace or matched GPU/FPGA comparison. These remain targets or hypotheses.

Bibliographic repair

Example ORCID fields and “final paper” headings do not establish publication. A cited “Kim and Jeong 2020” entry could not be verified as written. The verifiable Hu et al. dot-product-engine paper is in Advanced Materials, not Nature Electronics.

Patent scope

Specific restricted-write behaviour cannot be generalised to all IGZO devices. Authorisation and security tests are separate.

Confidential material

Confidential-agreement text and party information in the supplied files were not published. Commercial predictions without independent experiments are not stated as measured facts.