Is there something special about yolov8 over later models (9-12)? It seems most of the research and working examples default to v8 despite it being 3 years old. Or just because it is what fits on this hardware?
Mainly because YOLOv8 is well-supported by the Rockchip/RKNN toolchain.
The goal here was an end-to-end RK3588S pipeline rather than comparing detector families: training/export, ONNX graph fixing, INT8 RKNN conversion, C++ postprocessing, and runtime inference across the 3 NPU cores. YOLOv8 has known-good export paths and Rockchip examples, so it was the most practical baseline.
Newer YOLO versions may be possible, but usually require more work around RKNN export compatibility.
I built this while trying to understand how much of the RK3588S vision pipeline could be kept off the CPU.
The main trick is not the YOLO model itself, but the pipeline structure: MIPI capture through the ISP, resize/color conversion through RGA, and YOLOv8n inference through all 3 NPU cores with one RKNN context per core. With a 3-thread inference pool the pipeline goes from ~31 FPS to the OS08A10 camera’s 46 FPS ceiling.
The memory footprint is also small: roughly 137–152 MB RSS for one 1080p stream, using a fixed preallocated buffer pool rather than per-frame allocations. Two streams are roughly 276–304 MB RSS.
The repo also has a multi-process side of the pipeline: detections are published over Unix-domain sockets to tracking, temporal features, a presence FSM, and an optional Qwen2.5-0.5B summary step. For the LLM step, the camera pipeline can temporarily blackout/resume so RKLLM gets the whole NPU.
Batching is definitely the right answer for some offline / throughput-only cases, but it was not the right tradeoff here.
This pipeline is processing live camera frames and displaying/streaming annotated output, so latency and frame freshness matter. Increasing batch size would add queueing latency and tends to make the output older, especially when the sensor is producing frames continuously.
The “multithreading” here is not treating the NPU like a CPU in the usual sense. The RK3588S NPU is exposed as 3 cores, and RKNN supports using separate contexts with `rknn_dup_context` and assigning them with `rknn_set_core_mask`. The point was to keep the 3 NPU cores fed while capture, RGA preprocessing, inference, and display are pipelined.
In the single-context loop I was seeing ~31 FPS. With one context per NPU core and pipelined frame handling, it reaches the camera ceiling, around 42–46 FPS depending on the mode. So in this particular real-time streaming setup, parallel contexts/core masks were the practical way to saturate the hardware without adding batch latency.
Is there something special about yolov8 over later models (9-12)? It seems most of the research and working examples default to v8 despite it being 3 years old. Or just because it is what fits on this hardware?
Newer versions aren't open source, or at least have murky licencing.
Ahh that’ll do it. A shame really, the later models seem to be fairly good just from my idle testing as an enthusiast.
Mainly because YOLOv8 is well-supported by the Rockchip/RKNN toolchain.
The goal here was an end-to-end RK3588S pipeline rather than comparing detector families: training/export, ONNX graph fixing, INT8 RKNN conversion, C++ postprocessing, and runtime inference across the 3 NPU cores. YOLOv8 has known-good export paths and Rockchip examples, so it was the most practical baseline.
Newer YOLO versions may be possible, but usually require more work around RKNN export compatibility.
I built this while trying to understand how much of the RK3588S vision pipeline could be kept off the CPU.
The main trick is not the YOLO model itself, but the pipeline structure: MIPI capture through the ISP, resize/color conversion through RGA, and YOLOv8n inference through all 3 NPU cores with one RKNN context per core. With a 3-thread inference pool the pipeline goes from ~31 FPS to the OS08A10 camera’s 46 FPS ceiling.
The memory footprint is also small: roughly 137–152 MB RSS for one 1080p stream, using a fixed preallocated buffer pool rather than per-frame allocations. Two streams are roughly 276–304 MB RSS.
The repo also has a multi-process side of the pipeline: detections are published over Unix-domain sockets to tracking, temporal features, a presence FSM, and an optional Qwen2.5-0.5B summary step. For the LLM step, the camera pipeline can temporarily blackout/resume so RKLLM gets the whole NPU.
I split the work into three repos:
- runtime dual-stream YOLOv8n RK3588S pipeline: https://github.com/alebal123bal/khadas_yolov8n_multithread
- train/export/INT8 RKNN conversion for YOLOv8/YOLOv5: https://github.com/alebal123bal/RKNN_TRAIN_YOLO
- Qwen on RK3588S, via RKLLM/NPU or llama.cpp/CPU: https://github.com/alebal123bal/RKLLM_LLAMA_QWEN
The demo class is UAV/drone, but this is meant as a general edge-inference pipeline example, not an operational/surveillance/defense system.
These NPUs look very interesting.
Sad they are mostly sitting there unused because very few people know how to program them.
More slop again. The way to get more throughput is to bump batch size, not to try and "multithread" job submits to the NPU as if its a CPU.
Batching is definitely the right answer for some offline / throughput-only cases, but it was not the right tradeoff here.
This pipeline is processing live camera frames and displaying/streaming annotated output, so latency and frame freshness matter. Increasing batch size would add queueing latency and tends to make the output older, especially when the sensor is producing frames continuously.
The “multithreading” here is not treating the NPU like a CPU in the usual sense. The RK3588S NPU is exposed as 3 cores, and RKNN supports using separate contexts with `rknn_dup_context` and assigning them with `rknn_set_core_mask`. The point was to keep the 3 NPU cores fed while capture, RGA preprocessing, inference, and display are pipelined.
In the single-context loop I was seeing ~31 FPS. With one context per NPU core and pipelined frame handling, it reaches the camera ceiling, around 42–46 FPS depending on the mode. So in this particular real-time streaming setup, parallel contexts/core masks were the practical way to saturate the hardware without adding batch latency.
Again with it. You have two cameras, so you can batch 2 already with no latency hit. In fact less latency because the fake multithreading is gone.
(You are not even measuring latency correctly)