YOLO26n on RK3576: Built-in NPU vs RK182x vs Hailo-8

One model, three accelerator architectures

The reComputer RK3576 has a built-in NPU and one M.2 slot for an external accelerator. I ran the same YOLO26n INT8 workload on three options:

• RK3576 built-in NPU
• Hailo-8 M.2
• RK182x M.2

The comparison was split into two parts. First, I measured one video stream. Then I increased the number of independent streams to show how each architecture scales.

Test conditions

All three models used 640×640 input, COCO 80 classes, INT8 precision, the same six raw detection heads, byte-identical letterbox input, and one shared host implementation for decode and NMS.

The test clip contained 394 frames. A 640×640 derivative of the same 4K source was used because software decoding the original 4K file limited the host to 16.5 FPS.

Hailo-8 used Hailo's official HEF and recommended asynchronous pipeline, with four to eight frames in flight. The RK3576 and RK182x Python bindings were synchronous, so each platform was tested through its best available application path.

Single-stream result: Hailo-8 leads

For one stream, Hailo-8 was the clear winner. Its 49.2 FPS pipeline result was about 2.1 times the RK3576 NPU result, and its hardware latency was only 13.9 ms.

The final annotated-video rate was lower because drawing boxes and writing MP4 added roughly 15 ms of host work per frame. This is an application bottleneck, not accelerator compute.

Multi-stream result: RK182x scales highest

The multi-stream test used independent video decoding and inference for each stream. Drawing and video encoding were excluded so the result represents inference-service capacity.

The three architectures behaved very differently:

• Hailo-8 stayed near 51 FPS. One device served all streams, so adding streams divided its fixed throughput rather than increasing the total.
• The RK3576 NPU scaled to its available cores and reached 49.4 FPS with three streams.
• RK182x scaled across eight NPU cores and reached 88.6 FPS, 1.7 times the Hailo-8 aggregate result.

This is the main result of the project: RK182x is slower for one synchronous stream, but it is the strongest option when the workload contains many independent streams.

Conclusion

Hailo-8 is the strongest single-stream accelerator in this setup. The RK3576 NPU is a practical built-in option. RK182x becomes the best choice when four or more independent inference streams must run concurrently.

Reference

Three accelerators on one RK3576: built-in NPU vs Hailo-8 vs RK182x | reComputer AI Lab