EdgeFit Atlas

Measured on-device inference. Every number reproducible; every failure recorded.

Llama-3.2-1B-Instruct

hf:meta-llama/Llama-3.2-1B-Instruct · unknown

Graph fingerprint

1498.5M graph parameters (as exported) · 3838 graph nodes · attention: gqa

Constant×1353, Unsqueeze×496, Mul×263, Shape×231, Gather×198, Concat×196, MatMul×146, Cast×142

Time to first token by recipe

CPUaccelerator

Median of 10 timed runs after 3 discarded warmups. Hover a bar for variance and artifact size.

Repeatability

Recipes measured more than once on this unit, in separate sessions. This is the weakest useful form of the check — the real one runs the same recipe on two physical units of the same SKU, and we have one machine. Agreement here cannot catch a defect in the methodology that both runs share; it can only catch drift.

Recipe	sessions	fastest ms	slowest ms	spread
unknown	2	0.00	0.00	0.00%

Decode throughput

CPUaccelerator

Steady-state generation rate after prefill, with the KV cache threaded through every step. The chart above it is TTFT, which is the prefill cost — the two are separate workloads and are never averaged together.

Every recipe measured

Recipe	Target	Outcome	TTFT ms	tok/s	cv	size MiB	tokens	FLOP fb (auth)	time fb (run)	partitions
unknown	unknown	success	118.3	12.48	5.7%	5717	100%	—	0.0%	—
ort-cpu-dynamic	CPU	success	124.1	12.09	29.7%	5717	100%	—	0.0%	—
unknown	unknown	success	390.0	6.39	3.4%	5717	100%	—	15.0%	129
ort-coreml-dynamic	CoreML	success	473.0	5.26	20.4%	5717	100%	—	16.4%	129

Reproduce

Commands for every row on this page

# recipe 1fef51b8d3e705f5 is no longer in the library

uv run edgefit measure --model hf:meta-llama/Llama-3.2-1B-Instruct --recipe recipes/ort_cpu_dynamic.yaml

# recipe 229963667b8be77f is no longer in the library

uv run edgefit measure --model hf:meta-llama/Llama-3.2-1B-Instruct --recipe recipes/ort_coreml_dynamic.yaml