Compatibility registry

Which models run on which devices, on which accelerator, how fast, and with which fix. Every node benchmarks the models it runs and reports the result to its hub: works, works with tweaks (and the tweak), or fails (and why). The hub's router uses the same data.

11 results for 7 models on 2 devices (2 SoCs). So far they come from a Pixel 7 Pro and a TCL Google TV.

All results

Latency is the runner time on the accelerator (p50; no queueing or model load) unless the note says otherwise. Filters are kept in the URL, so a filtered view can be shared.

Showing the 2026-11-18 snapshot, checking the live registry…

11 of 11 results

Model and device compatibility results. Column headers sort the table.
clip-l14-336
image-embedding

~1450 ms per image (two views: letterbox + centre crop); fp16 weights, 582 MB; 1.65 s round trip from the Mac

Google Pixel 7 Pro
Google Tensor G2
GPU
litert-gpu
Works1.45 s
efficientnet-lite0-int8
image-classification

invoke time

TCL Smart TV Pro
MediaTek MT9689
NPU
nnapi:mtk-mdla_shim
Works6 ms
efficientnet-lite2-int8
image-classification

invoke time

TCL Smart TV Pro
MediaTek MT9689
NPU
nnapi:mtk-mdla_shim
Works9.4 ms
mobilenet-v2-int8
image-classification
Google Pixel 7 Pro
Google Tensor G2
CPU
cpu
Works7 ms
mobilenet-v2-int8
image-classification
Google Pixel 7 Pro
Google Tensor G2
NPU
nnapi
Works0.78 ms
mobilenet-v2-int8
image-classification

4.4–5.0 ms invoke

TCL Smart TV Pro
MediaTek MT9689
NPU
nnapi:mtk-mdla_shim
Works4.4 ms
p95 5 ms
efficientnet-lite4-int8
image-classification

Tweak: Load the model from a memory-mapped buffer, not a path on shared storage: SELinux blocks the NN HAL (mtk_hal_neuralnetworks) from reading FUSE files, and NNAPI then silently falls back to its CPU reference path (~9 s per image).

TCL Smart TV Pro
MediaTek MT9689
NPU
nnapi:mtk-mdla_shim
Works with tweaks22 ms
mobilenet-v2-fp32
image-classification

Tweak: Enable allow_fp16 on the NNAPI delegate (setAllowFp16(true)).

TCL Smart TV Pro
MediaTek MT9689
NPU
nnapi:mtk-mdla_shim
Works with tweaks11.2 ms
clip-vit-b16
image-embedding

fp32, 330 MB: out of memory while loading; the TV rebooted.

TCL Smart TV Pro
MediaTek MT9689
Unknown
n/a
Fails–
clip-vit-b16
image-embedding

The GPU delegate does not support the graph (unsupported CONCATENATION).

TCL Smart TV Pro
MediaTek MT9689
GPU
gpu (Mali-G52 delegate)
Fails–
mobilenet-v2-fp32
image-classification

The MDLA does not run fp32 graphs; without allow_fp16 NNAPI rejects the model for the NPU.

TCL Smart TV Pro
MediaTek MT9689
NPU
nnapi:mtk-mdla_shim
Fails–

Model × device matrix

Best result per pair, from the 2026-11-18 snapshot.

Best result for each model on each device
ModelTCL Smart TV ProGoogle Pixel 7 Pro
clip-l14-336not testedWorks1.45 s
clip-vit-b16Fails–not tested
efficientnet-lite0-int8Works6 msnot tested
efficientnet-lite2-int8Works9.4 msnot tested
efficientnet-lite4-int8Works with tweaks22 msnot tested
mobilenet-v2-fp32Works with tweaks11.2 msnot tested
mobilenet-v2-int8Works4.4 msWorks0.78 ms

How results get here

After a model's first five runs on a node, the node sends p50/p95 runner times and a status to its hub: works (loaded on the manifest's first-choice accelerator), works with tweaks (it fell back; the note says why) or fails (the load error). The owner's measured TV and Pixel results seed the registry. Public submissions from other hubs are planned.

How the registry worksThe benchmark frame