Home / Engines / Model Match

Which model should run on the hardware you already have?

One platform in, one model out. Every model with a benchmark row for that silicon is a candidate — after a canonical map joins the two ids the benchmark database records the same checkpoint under. Each is evaluated once for throughput, latency, precision, runtime and batch, and once for memory, then filtered on hard facts only and ranked by one published weight table. Accuracy is reported in two layers that are never merged: the model's own published figure with the metric and the dataset named, and the quantised figure observed at your precision. The winner is the top of the list — there is no override.

01 · Describe the question

Platform & question
Platform
Loading platform catalog…
Task
Optimise for
Balanced
Accuracy
Speed
Size
Precision
The load
Load from
Cameras
A frame-rate target
Cameras
1
4
8
16
32
Detect rate
1 fps
5 fps
10 fps
15 fps
30 fps
Frames with motion
20 %
50 %
80 %
All (worst case)
Latency target
Not stated
15 ms
33 ms
100 ms
500 ms
Advanced
Runtime
Model input
640
320
416
1024
Batch size
1
2
4
8
Headroom to keep
None
10 %
20 %
40 %
Power mode
Models to evaluate
10
20
40
All (90)
Quantisation language models
Q4
INT8
FP16
Context language models
2 k
4 k
8 k
32 k
Concurrent sequences language models
1
2
4
8
Loading the platform catalog…

03 · How this works

The candidate universe is the models measured on your silicon

A model is a candidate here only if somebody has published a benchmark row for it on the platform you named. That is a deliberately narrow rule. The alternative — admitting every model and modelling the ones nobody measured — fills the table with rows that can never support a choice, because the number that would decide between two models is the one that was invented. A model with no row on your platform is reported as a gap, never as a silently modelled entry, and the result says how many rows the platform has at all.

Before the universe is built, a canonical map joins the ids the benchmark database records the same checkpoint under. On the Hailo-8L, SSD MobileNet v1 has its throughput rows under one id and its accuracy row under another; the same split exists for the segmentation and pose variants of YOLOv8 and YOLO11, and for three different spellings of the EfficientNet-EdgeTPU family. Without the join, one candidate would have speed and no accuracy while a phantom second had accuracy and no speed.

Nothing is recomputed — every figure comes from the engine that owns it

Throughput, latency, precision, runtime and batch are Compute Fit's answers, unchanged: the pessimistic edge of the matching benchmark rows, the three latency tiers that are never blurred, and UNKNOWN where nothing resolves. Memory for a language model is Model Memory Fit's. Memory for a vision model is the weight file plus the published TensorRT activation and workspace split, on the shared memory budget — and where no split was published, that curve is interpolated on to an architecture it was not measured on and labelled class E for exactly that reason.

One filter, one ranking, no override

A candidate is excluded if and only if a compute check fails or its memory does not fit. Not on accuracy, not on size, not on a preference: a preference changes the order of the ranked list, never its membership, and every excluded model comes back with the constraint that excluded it. UNKNOWN never excludes anything — it caps the verdict at NEEDS VALIDATION and pulls the evidence term down.

The survivors are ranked once, on four terms in [0,1]: accuracy (this model over the best in its own task and metric cohort, and zero when no accuracy was ever published), margin (one minus the throughput utilization), evidence (the mean class weight of the checks and the accuracy fact) and size (the smallest weight file over this one). The weight table is class E engineering judgement and is printed beside the scores so the arithmetic can be checked. The winner is the top of that list, full stop.

Full rules, the weight table and what invalidates a result: methodology. Related: Compute Fit answers for one named model; Hardware Match answers which platform; Benchmark Explorer is the row database underneath.

04 · FAQ

Why are there two accuracy numbers for the same model?
Because they are two different measurements and merging them would hide that. The reference accuracy is the figure the model’s publisher puts on its own model card, with the metric and the evaluation set named — for detectors that is mAP50-95 on COCO val2017. The quantised accuracy is what somebody observed at the precision you asked for, on the platform you asked about. For YOLO11n those two are 0.395 and 0.5085, eleven points apart, which is far more than quantisation can explain: they cannot be the same metric on the same set, and the second source does not say which set it used. So both are shown, the gap is stated, and the ranking uses the published reference figure.
Why does a model with no published accuracy score zero rather than being left out?
Because leaving it out would be a silent judgement and giving it credit would be a fabricated one. Roughly half the models with benchmark rows on edge silicon have no accuracy figure anybody published. Such a model still competes on throughput headroom, on the evidence class of its numbers and on size, and it can still win — but it never borrows accuracy it has not earned. The result says how many candidates are in that position.
Can I make it pick a different model?
Not directly, and that is the point. The winner is the top of the ranked list, computed by one weight table that is published in full beside the scores. If you disagree with the answer, change the question: pick a different preference, tighten the latency target, narrow the precision or exclude a family. There is no override, no re-pick and no hook that lets a caller splice a favourite to the front — a ranking with an override is not a ranking.
Why is a model missing from the list?
Either nobody has benchmarked it on this platform, or it was excluded on a hard fact. A model with no benchmark row for this silicon is not a candidate at all: it would have to be modelled from another platform, and a modelled figure cannot support a choice between models. A model that was evaluated and then excluded appears in the excluded list with the exact constraint that excluded it — a compute check that failed, a precision the silicon does not support, or a memory that does not fit.
What is the canonical model map for?
The benchmark database records the same checkpoint under two names. SSD MobileNet v1 on the Hailo-8L has its throughput rows under one id and its accuracy row under another; YOLOv8n-seg has seven throughput rows on one id and its 0.367 mAP on a second. Without a map that joins them, one candidate would have speed and no accuracy while a phantom second had accuracy and no speed. The map is published in the model-facts dataset, every entry carries the reason it exists, and the result says how many of the candidates it joined.