Which platform, at how many nodes, and why that one?

Describe the workload and the constraints that are real. Every platform in the catalog is then evaluated at every node count from one to the cap, through the engines that own each fact — decode and inference through Camera Stream Capacity, memory through Model Memory Fit, thermal through the Thermal Feasibility Checker, JetPack and runtime through the Jetson Configuration Checker, carriers, the module power ceiling and the PoE path through their own engines. A candidate is eliminated only on a hard fact, never on a soft penalty and never on budget. The survivors are ranked once, by one published weight table, on five terms: margin, evidence, cost, power and simplicity. The winner is the top of that list — there is no second pass and no re-pick.

01 · Describe the workload

Cameras & streams
Cameras
4
8
16
32
48
Resolution
720p
1080p
1440p
4K
Frame rate
10 fps
15 fps
25 fps
30 fps
Codec
H.264
H.265
Task
Detection
NVR (Frigate)
Precision
INT8
FP16
FP32
Constraints
Optimise for
Balanced
Lowest cost
Lowest power
Most headroom
Power ceiling
None
60 W
150 W
400 W
Cooling
Fan allowed
Fanless
Ambient
Not stated
25 °C
40 °C
55 °C
Budget
Not stated
Low
Moderate
High
Node power
DC supply
PoE
Advanced
Node cap
1
2
4
8
Restrict to
Loading platform catalog…
JetPack
Install method
Not stated
SDK Manager
ISO image
apt upgrade
Carrier needs
2× GbE
2× CSI
Industrial temp
Loading the platform catalog…

03 · How this works

One filter, one ranking function, one winner

The filter is deterministic and uses hard facts only. A candidate is infeasible when a constraint comes back FAIL or UNSUPPORTED at every node count up to the cap, or when the node count times the unit power exceeds a ceiling you stated. There is no × 1.5 tolerance on stream counts, no soft penalty for a platform in preview, and no multiplier applied to an interpolated benchmark. Budget is not part of the filter: a platform you cannot afford is a different problem from one that cannot do the job, so over-budget candidates stay in the list and are labelled.

The survivors are ranked once, by Σ wᵢ sᵢ over five terms, each in [0, 1]. Margin is one minus the worst utilization across the candidate's constraints. Evidence is the mean evidence class of those constraints, where an UNKNOWN counts as the weakest class. Cost and power are the cheapest and lowest totals divided by this candidate's totals. Simplicity is the fewest nodes divided by this candidate's node count. Nothing else enters the score — no ecosystem rating, no SDK-maturity opinion, no goal-specific branch. The weights are published, and you can recompute any score from the numbers the result returns.

Node count is part of the candidate

Each platform is evaluated at one node, then two, and so on to the cap; the reported node count is the smallest that fits. Cameras are split by a balanced assignment, so sixteen across six nodes is 3 / 3 / 3 / 3 / 2 / 2 and sums to exactly sixteen. Cost and power are always the node count times the unit figures, so a cheap module that needs eight nodes competes honestly against an expensive one that needs two. Because node count is chosen inside the ranking rather than corrected afterwards, there is no economics winner to splice to the front of the list.

UNKNOWN is never fatal and never free

A missing registry row does not eliminate a platform — only a real failure does. It does cap that platform's verdict at NEEDS VALIDATION and pull its evidence term down, and a candidate whose decode, memory and inference are all unknown is listed after every candidate with coverage, so it cannot win while a checkable option exists.

Full rules, the class E weight table and what invalidates a result: methodology. Related: Camera Stream Capacity is where the decode and inference numbers come from; Thermal Feasibility Checker owns the enclosure question; Jetson Power & PSU sizes the supply once the platform is chosen; Carrier Finder finds the board it plugs into.

04 · FAQ

Why does the winner differ from the Hardware Selector?
Because five layers picked the platform there and one picks it here. The legacy path ran a goal-weighted rank, then spliced a multi-node economics winner to the front, then let the orchestrator re-pick on a confidence tier. It also weighted an ecosystem score of SDK maturity, model breadth and documentation quality that has no source behind it. Hardware Match drops the ecosystem term entirely, has no re-pick and no splice, and makes node count part of the candidate rather than a correction applied afterwards.
Why is my budget not removing the expensive platforms?
Budget is deliberately not a feasibility filter. A platform that costs more than you planned is a different decision from one that cannot do the job, and collapsing the two hides the option you might actually take. Over-budget candidates stay in the ranked list, flagged, and can carry the stretch role. What does remove a candidate is a hard fact: a failing constraint, or a node power above a ceiling you stated.
What does an UNKNOWN check do to the answer?
It never eliminates a candidate and it is never free. An UNKNOWN caps that candidate's verdict at NEEDS VALIDATION and counts as class E in the evidence term, so a platform with real datasheet coverage outranks one we simply cannot check. A candidate whose decode, memory and inference are all unknown is listed after every candidate with coverage, so it can never be the winner while a checkable option exists.
Where do the ranking weights come from?
They are engineering judgement, class E, and they are published in full — on the methodology page, in the API catalog and on every result. There is exactly one table with one row per goal; no overlay, multiplier or goal-specific branch exists anywhere else in the engine. You can recompute any score from the per-term numbers the result returns.
Why are the accelerator prices lower than the nodes they build?
A Coral stick or a Hailo board is a card, not a node: it needs a host to decode, pre-process and run the OS. The host's power is included in the figures — it is a published vendor maximum — but no vendor listing for the host is on record, so the price shown is the card alone and the cost evidence drops to class E. No host price is invented to fill the gap.