02 // INTERACTIVE TELEMETRY LAB

Real multi-stream RKNN INT8 benchmarks.

Select a computer vision model and concurrent 1080p RTSP stream count below to inspect real hardware telemetry measured on our fanless ArkBox appliances at 28°C ambient room temperature.

BENCHMARK TELEMETRY // LAB NODE: ARKBOX-88 (RK3588)

rknn-toolkit2 v2.3.0 · HYBRID INT8

1. Select Vision Workload

2. Concurrent 1080p RTSP Camera Streams

INFERENCE SPEED / CH

31.4 FPS

Total throughput: 251.2 FPS

NPU UTILIZATION

78% 3-Core NPU

MPP HW DECODE CPU

18% Cortex-A76/A55

Zero CPU soft-decode bottleneck

WALL POWER & TEMP

11.8W / 51°C

Passive fanless @ 28°C ambient

root@arkbox-88:~/inferark-bench# ./run_rknn_mpp_pipeline.sh FP32 mAP: 54.2% → INT8 mAP: 53.8% (-0.4% loss)

[MPP-DEC] Initialized 8x H.264 1920x1080@30fps RTSP hardware decoder queues
[RKNN-NPU] Loaded yolov8s_ppe_hybrid_int8.rknn across NPU Core 0 | Core 1 | Core 2
[GPIO/RS485] /dev/ttyS4 Modbus relay ready ( Trigger latency < 8ms )
[STREAM 01..08] Avg latency: 11.9ms | Per-channel: 31.4 FPS | Dropped frames: 0

TURNKEY SOFTWARE ARCHITECTURE

Why bare boards stall your team—and what we pre-flash.

Every InferArk appliance ships with a production-grade Ubuntu 22.04 LTS + Docker image featuring zero-copy DMA buffer sharing between the Rockchip hardware video decoder and the 3-core NPU.

03 // TENDER BOM CALCULATOR

Quantify your hardware savings on multi-node rollouts.

Slide to select the number of edge appliances in your upcoming UAE, Malaysia, or European deployment. Pricing automatically adjusts across our 1–9, 10–49, and 50+ unit FOB tiers.

Deployment Volume (Edge Boxes) 25 Units

5 units 50 units 100 units 200 units

INDUSTRIAL JETSON ORIN NANO BOX (AVG $890 / UNIT)

Bare hardware · Self-managed TensorRT porting

$22,250

INFERARK ARKBOX-88 (FOB $295 / UNIT)

Includes free ONNX→RKNN INT8 porting + pre-flashed Docker OS

$7,375