Hardware Selector — methodology

Method version 4.0.1 · engine: /engines/hardware-selector/ · runs Hardware Match and reads the production-carrier registry (IP67 rows only).

The Selector is a simple front end for Hardware Match. Its answer is Hardware Match’s answer to the request its tiles describe; the hand-off link opens Hardware Match with that request and every control it has.

Contents

  1. Purpose, and its relationship to Hardware Match
  2. Inputs and how each one reaches Hardware Match
  3. The answer
  4. Refusals and suggestions
  5. The latency requirement
  6. Compliance
  7. Scenario and ambient temperature
  8. Evidence limitations
  9. Method changelog

1. Purpose, and its relationship to Hardware Match

The Hardware Selector answers one question quickly: for a camera workload described in a few tiles, which platform, at how many nodes? Since method 4.0.0 it does not size, filter or rank anything itself. It turns its inputs into one Hardware Match request, runs it, and presents the answer. The recommendation is Hardware Match’s winner on that request, at the same node count, and the Selector refuses exactly when Hardware Match has no recommendation-eligible winner. How the platform and node count are decided is on the Hardware Match methodology page.

Every result carries hardware_match (its method version, verdict and winner) and scope_disclosure: the exact request as hardware_match_inputs, a hand-off link that opens Hardware Match with it, and not_carried — the Selector inputs that were set but that Hardware Match is not asked about, each with a sentence in differs.

2. Inputs and how each one reaches Hardware Match

InputPage valuesHardware Match request
Task (model)classification, object detection, segmentation, pose, multi-modeltask: detection (object detection, real-time detection, tracking and multi-model), classification, segmentation, pose. For a multi-model pipeline only the detector is sized; Architecture Designer sizes the whole pipeline. Generative and language tasks are refused with OUT_OF_SCOPE and a link to Model Match.
Precision(API only)Sent when named. With no precision named, the task’s planning default applies: FP16 for segmentation and pose, INT8 for every other task. precision is listed in defaulted.
Streams1–16cameras, at 15 fps (a stated fps is named in not_carried). Hardware Match plans up to 8 nodes.
Resolution720p, 1080p, 4Kresolution
Power≤ 30 W, ≤ 80 W, ≤ 300 W per device (≤ 10 W by link and API: no complete platform fits it, so the page does not offer it)node_power_limit_w 10 / 30 / 80 / 300: a platform whose maximum module power, or whose accelerator card plus host, exceeds it is excluded at every node count. unlimited (API) sends no ceiling. When a PoE standard powers the node (poe_standard, API), the power that standard delivers to the powered device is the ceiling if it is lower (802.3af 12.95 W, 802.3at 25.5 W, 802.3bt 51 / 71.3 W).
Environmentfanless, active cooled, industrial, mobile robot, edge rackFanless sends fanless: true to Hardware Match’s thermal check. Industrial (outdoor exposed) is answered for a vented enclosure and named in not_carried (section 6). The other values are not carried.
Budget sensitivitylowest cost, balanced, highest performanceoptimize_for: cost, balanced, headroom.
Latencybatch, near real-time, real-time criticalNot sent; reported on the winner (section 5).
Compliancenone, IP67Not sent; sourced IP67 rows on the winner (section 6).
Scenariosix deployment patternsNot sent; labels the summary (section 7).
Ambient temperature25 / 35 / 45 / 55 °Cambient_max_c, judged by Hardware Match’s thermal check (section 7).

Inputs (WP-01). A field that is not stated takes its documented default and is listed in provenance.defaulted; a field that is stated but invalid — a count of 0, a negative, a fraction or a non-number, true in place of a count, an unknown value of an enumerated input, or an empty string — is rejected with VALIDATION_ERROR naming the field, never replaced by a default.

Input keys (WP-02). The engine accepts one list of input keys — the inputs documented here and their declared backward-compatible aliases. Unknown input keys are rejected with VALIDATION_ERROR naming the key: an invented or misspelt key is never silently ignored, a __ test hook never reaches the engine, and an alias sent together with the key it stands for is rejected too (the engine would read only one of them). export_json.inputs echoes every input the engine honoured, so replaying it reproduces the result.

3. The answer

The recommended device, node count, per-node stream assignment, unit and total cost, confidence label and limiting factor are Hardware Match’s, for its winner. The per-node figure is the streams Hardware Match validated on the busiest node, not a maximum; the headroom shown is Hardware Match’s modelled utilisation of that node’s binding limit (binding_utilization_pct). No stream capacity is extrapolated beyond what Hardware Match checked, because its limits are not linear in cameras. Deployment prerequisites Hardware Match names (for example a flash configuration) are carried as deployment_prerequisites; a fanless request reports the cooling Hardware Match’s thermal check passed, not the module’s reference heatsink. Alternatives are Hardware Match’s next recommendation-eligible configurations, up to four, labelled from Hardware Match’s roles (lower cost, lower power, more headroom) or as needing fewer nodes. Compute, power range and cooling come from the platform catalog.

4. Refusals and suggestions

When Hardware Match has no recommendation-eligible winner, the Selector answers NO_ELIGIBLE_PLATFORMS. blocked_by groups the platforms Hardware Match excluded by the constraint that excluded them (power, precision, cooling, ambient, decode, memory, compute, streams), with examples. planning_only_candidates lists the configurations Hardware Match kept for planning only, best first, with Hardware Match’s reason (capacity modelled rather than measured, heuristic evidence, or unresolved constraints); at most eight are listed and planning_only_total gives the count. When a stated precision leaves every candidate for planning, the message says no platform has measured capacity at that precision. nearest_fits re-runs the Selector with one input relaxed at a time (power, environment, streams, resolution, ambient, compliance), at most 20 runs, and lists only the changes that produce a recommendation.

5. The latency requirement

Targets (planning definitions, class E): real-time critical 33.3 ms per frame, one frame interval at 30 fps; near real-time 100 ms; batch no target, not evaluated. With no latency input the requirement is not evaluated.

The winner’s modelled per-frame latency, and that of each alternative, is 1000 divided by the single-stream frame rate at batch 1 that Hardware Match modelled for its inference constraint. It is class D: it excludes capture, network and queueing and is not a measured end-to-end latency. The requirement is informational: it is compared with the target and reported in latency_check (PASS, FAIL, UNKNOWN when no frame rate is modelled, NOT APPLICABLE with no target), and it never changes the recommendation. Relaxing or tightening the requirement therefore returns the same configuration.

6. Compliance

IP67 is a sourced check against the production-carrier registry: a platform is PASS when a registry system for its module carries an IP67 rating that a cited source states, and UNKNOWN when no such system is on record. UNKNOWN is not PASS. The registry records no module as lacking an IP67 option, so the check flags and never removes a configuration; compliance_checks publishes one row for the winner and each alternative, with the evidence behind each PASS. IP67 and the industrial (outdoor exposed) environment are not sent as Hardware Match’s sealed outdoor enclosure: Hardware Match has no thermal data for a module inside one and holds every module for planning only there. The Selector answers for a vented enclosure, names sealed_enclosure in not_carried, and the hand-off opens Hardware Match with the sealed enclosure, where the thermal check stays open.

ATEX and MIL-STD-810 are not modelled: no platform, carrier or enclosure in the registries carries a sourced ATEX or MIL-STD-810 certification. The page no longer offers them, and a request that names either is refused with a validation error on compliance_requirements rather than answered as if the requirement had been checked. Confirm those certifications with the enclosure vendor.

The advisory compliance values the API accepts (NDAA, EU AI Act, HIPAA, NERC CIP, ISO 26262, PCI DSS) produce warnings only; an unknown value is a validation error.

7. Scenario and ambient temperature

Scenario is informational. It labels the summary and flags capability-profile caveats — for example an accelerator whose profile says it is weak for this deployment pattern — and input_effects.deployment_scenario reports the winner’s profile fit. It does not change the ranking, the capacity or the cost; it is not sent to Hardware Match.

Ambient temperature is sent to Hardware Match as the site’s maximum ambient. Hardware Match’s thermal check excludes a candidate whose sourced thermal limits cannot hold it, so a hot site can change the recommendation or produce a refusal grouped under ambient.

8. Evidence limitations

  • Everything Hardware Match’s methodology lists as a limitation applies, because the answer is Hardware Match’s.
  • The Selector fixes the frame rate at 15 fps and does not carry the scenario, the latency requirement, flexibility, future growth, topology or reliability; open the hand-off to set what Hardware Match offers.
  • Hardware Match does not yet apply the precision-support registry to a vision precision: an INT8-only accelerator asked for FP16 is held for planning rather than excluded (#289). It is never recommended there, because no measured FP16 capacity exists for it.
  • A sealed outdoor enclosure is not modelled in the Selector’s answer (section 6).
  • Modelled latency is class D; the real figure depends on the model, runtime and pipeline — measure on the target.
  • The IP67 check reads one registry and covers only systems recorded there; an absent rating means unknown, not unrated.

9. Method changelog

MethodDateChange
4.0.12026-11-07A refusal caused by a stated precision the vendor documents as unsupported is grouped as “Precision not supported (vendor-documented)” (Hardware Match 1.23.0, #289). The page no longer offers the ≤ 10 W tile, which no complete platform fits; power=ultra_low still works by link and API.
4.0.02026-11-07Breaking: the Selector answers from Hardware Match (owner ruling 2026-11-06). Its own hard filters HF2–HF11, capacity model, complete-configuration ranking and memory fit are removed; the recommendation is Hardware Match’s winner on the mapped request (section 2), and a refusal is built from Hardware Match’s exclusions and planning-only configurations. Power classes become per-device ceilings sent as node_power_limit_w (10 / 30 / 80 / 300 W, the limits the old filter applied; the page labels said 5 / 10 / 25 / 25+ W). Segmentation and pose with no precision follow the FP16 task default (the INT8 hold is removed). More than 32 cameras is no longer refused; generative and language tasks are refused with OUT_OF_SCOPE. The latency requirement no longer reorders configurations. On a 5,184-input grid, refusals fall from 34.7 % to 26.6 % (12.9 % to 2.2 % outside the ≤ 10 W class, which refuses before and after).
3.1.02026-09-29Inline Selector platforms Hailo-10H, RK3588S, RK3568 and Neousys Nuvo-9531 now read supportedPrecisions from the same precision-support registry as catalog platforms at module load (bug #277), so HF6 removes them for vendor-documented unsupported precisions (for example Hailo-10H at FP16, Rockchip SBCs at FP32 or INT4). OnLogic Factor 201 has no registry row and remains without a list. Winner changes only where those platforms had incorrectly stayed eligible under an unsupported precision.
3.0.42026-09-29The precision filter HF6 reads the vendor-sourced support lists (bug #275): Jetson Thor T5000 is no longer removed for FP8, BF16 or INT4, the Jetson Orin modules and Hailo-8 / 8L for INT4, and RK3588 is now removed for INT4 (Rockchip's toolkits deploy no 4-bit weights on it); each removal records whether a vendor documents the exclusion (class A) or it is a catalog assertion (class E). The assumption that states the segmentation / pose hold is now a plain sentence without a ledger reference. No recommendation changes on the reference workloads.
3.0.32026-09-29An absent precision is listed in defaulted, with an assumption naming the default. Segmentation and pose stay at INT8: the shared FP16 task default (owner ruling of 2026-09-29) is held at INT8 for the Selector, #274, until its per-node capacity comes from one model — with the FP16 default its plans were not monotone in cameras. Every recommendation is unchanged.
3.0.22026-09-28Input rules (WP-01 remainder, WP-02): power / power_sensitivity, env / environment, scenario, model, resolution, flexibility_requirement and budget_sensitivity are checked and the error names the key sent; ambient_temperature_c "abc" is rejected instead of becoming 25 °C; defaults are listed in provenance.defaulted. Unknown input keys are rejected. No formula changed.
3.0.12026-09-28Present-invalid inputs are rejected; absent inputs default and are recorded (WP-01): a stated but malformed streams / stream_count or camera_count — 0, negative, "four", true, 2.5, "" or above 256 — returns VALIDATION_ERROR naming the field instead of becoming the 4-stream default (or eliminating every platform); a stream count that is not stated takes the 4-stream default and is listed in provenance.defaulted. No formula changed.
3.02026-09-28Breaking: the ATEX and MIL-STD-810 compliance values are removed (no sourced certification facts) and now return a validation error on compliance_requirements, as does any unknown compliance value or latency requirement. The latency requirement is wired as an ordered class D ranking preference with latency_check rows (real-time critical 33.3 ms, near real-time 100 ms, batch not evaluated). IP67 is a sourced check against the production-carrier registry with compliance_checks rows (PASS / UNKNOWN, never removes). Scenario and ambient temperature are published as informational and heuristic in input_effects. The result states it is an estimate and hands Hardware Match the exact inputs with the pick pinned (hardware_match_inputs, not_carried). First methodology page.
2.32026-09-27Last version before this page. The latency, compliance (ATEX, IP67, MIL-STD-810) and scenario controls were accepted and changed nothing in the public result.

Method changes bump the method version; results carry it in provenance.method_version with the dataset version of the registries the Selector reads.