For TV, set-top box and router makers

The node can be built into a device. It joins the household’s hub over an outbound connection, serves models on the device’s NPU and reports what runs well on that chip.

Measured on a TCL Google TV

A retail TCL Smart TV Pro (MediaTek MT9689, 2.4 GB RAM) running the Android app, measured by us in November 2026. Not an endorsement by the manufacturer.

Modelp50Notes
EfficientNet-Lite4 int822 msInvoke time on the MDLA NPU via NNAPI. 118 of 120 graph nodes run on the NPU; the two on the CPU are the input and output QUANTIZE ops (under 0.7 ms together).
MobileNetV2 int84.4 msSame NPU.
MobileNetV2 fp3211.2 msOnly with fp16 relaxation enabled on the NNAPI delegate; without it the NPU rejects the fp32 graph. The registry records this as a tweak.
CLIP ViT-B/16 fp32fails330 MB: out of memory while loading, and the TV rebooted. The GPU delegate does not support the graph either.

What we learned on real hardware

Storage and SELinux
Vendor NN HALs may not read app files on FUSE shared storage; NNAPI then silently falls back to a CPU path about 400× slower. Load models from memory-mapped buffers.
Power states
The tested TV suspends about 3 minutes after power-off even with a wake lock. Low-power standby policies decide whether a device can be an always-on node.
Memory headroom
About 700 MB was free on a 2.4 GB TV. A 330 MB fp32 CLIP ran it out of memory; small int8 vision models fit comfortably.
Footprint
The Android node APK is about 7.7 MB for the 32-bit NNAPI build. A per-model memory budget keeps it within what the device can spare.

The node only dials out to the household’s hub over TLS, so it adds no listening port to the device. Registry data shows which models run on your SoC, on which accelerator, how fast and with which tweaks. See the registry.

Building the node into your devices

The software is free. Terms for shipping the node inside your own devices are not set yet; tell us about your device line and we will get in touch.