Optimized builds

Compare reference and optimized performance on the same board, with precision, quality results and measurement conditions for each run.

Workload

Board

Baseline

5 models · 3 boards · 7 ZETIC comparisons in view

ZETIC optimization overview

Reference vs optimized

Latency on the same model and board.

Lower is better

Latency · Log scale

Selected baseline and fastest recorded ZETIC build on the same board. Test conditions in Details.

Latency · p50Baseline / ZETIC · speedup on the right

Baseline and ZETIC latency on the same board, grouped by model. Quality results retain their evaluation context. Open Details for key configuration and test conditions.
Model	Board	Latency	Improvement	Quality result
SmolVLA	Orin Nano	Baseline 1015.3 ms ZETIC 85.76 ms per action chunk · p50	11.8×faster	libero spatial success rate: 0.86 LIBERO Spatial · Quality reported · no pass/fail verdict
	AGX Thor	Baseline 197.21 ms ZETIC 29.72 ms per policy call · p50	6.6×faster	max abs: 2.011076867580414 LIBERO Spatial · Quality reported · no pass/fail verdict
	AGX Orin	Baseline 750.895 ms ZETIC 35.39 ms per action chunk · p50	21.2×faster	max abs: 2.072598695755005 LIBERO Spatial · Quality reported · no pass/fail verdict
GR00T N1.7	AGX Thor	Baseline 108.52 ms ZETIC 23.46 ms per policy call · p50	4.6×faster	Median absolute error: 0.00339 LIBERO 10 · Quality reported · no pass/fail verdict
Cosmos3 Edge Policy DROID	AGX Thor	Baseline 726.05 ms ZETIC 552.62 ms per policy call · p50	1.3×faster	max abs: 0.5365428924560547 DROID · Quality reported · no pass/fail verdict
RLDX-1	AGX Orin	Baseline 607.61 ms ZETIC 188.75 ms per action chunk · p50	3.2×faster	max abs: 0.5168914794921875 LIBERO Object · Quality reported · no pass/fail verdict
GR00T N1.6 G1 PnP	AGX Orin	Baseline 308.66 ms ZETIC 91.48 ms per policy call · p50	3.4×faster	success rate: 0.6 Unitree G1 apple to plate · Quality reported · no pass/fail verdict

Loading benchmark filters…

Build with ZETIC

Share your model, hardware, and goals for latency, energy per inference, and memory.

We’ll agree on scope, then benchmark, optimize, and validate against your baseline and quality targets.