Edge AI solution provider, Aetina Corporation, has launched the AIE-KR1B-B1, an ARM-based edge AI system powered by the Rockchip RK3588 8-core 64-bit processor.
In addition to the RK3588’s integrated NPU, the AIE-KR1B-B1 provides a PCIe Gen3 x4 M.2 expansion interface for discrete AI accelerators, allowing customers to select accelerator options based on workload, performance, power, and cost requirements without redesigning the host hardware.
A heterogeneous architecture for flexible edge AI acceleration
Edge AI workloads involve more than AI inference alone. In multi-stream vision applications, the host platform must also handle video ingestion, decoding, image processing, data preparation, post-processing and system I/O before and after inference.
The AIE-KR1B-B1 uses a heterogeneous computing architecture to divide these workloads across dedicated processing resources. The Rockchip RK3588 serves as the host processor, handling the host-side vision pipeline, while the primary AI inference workload can be offloaded to a dedicated M.2 AI accelerator. This allows the RK3588 to focus on video and system processing while the accelerator focuses on AI inference.
The AIE-KR1B-B1 is part of Aetina’s AI Vision Solutions, enabling flexible integration of M.2 AI accelerators through its PCIe Gen3 x4 M.2 interface. It supports AI accelerator options from Axelera AI, DEEPX, EdgeCortix, and MemryX, spanning a compute range of 25 to 214 TOPS based on vendor specifications. Customers can select an accelerator according to AI model, performance, power and cost requirements while maintaining the same RK3588-based host platform.
Beyond the primary AI M.2 slot, the platform incorporates an additional M.2 M-Key interface for SATA storage, an M.2 E-Key slot for Wi-Fi/Bluetooth, and a Mini PCIe slot for LTE, alongside dual Gigabit Ethernet, USB 3.2 Gen1, USB 2.0, configurable RS-232/RS-485 COM, and HDMI 2.0 — supporting both wired and wireless edge deployments.
Validated Multi-Stream Video Performance Ahead of Accelerator Selection
For multi-camera edge AI applications, overall system performance depends not only on AI inference, but also on how efficiently the host platform receives, decodes, converts, and prepares video data before it reaches the AI accelerator.
To provide customers with greater visibility into host-side video-processing capability, Aetina validated the RK3588 video pipeline on the AIE-KR1B-B1 using up to 32 concurrent 1080p video streams.
The test measured host-side processing latency across multiple stream counts, including video decode, color conversion, compositing and additional processing stages before AI inference. The test measured the complete host-side processing flow across different numbers of simultaneous video streams:
| Pipeline Stage | 16ch | 20ch | 25ch | 32ch |
|---|---|---|---|---|
| Decode (VPU, 1080p) | 111.7 ms | 143.6 ms | 174.8 ms | 225.2 ms |
| Color Convert (CPU, NV12 → BGRx) | 2.9 ms | 2.2 ms | 1.9 ms | 1.5 ms |
| Composite | 60.0 ms | 61.0 ms | 63.0 ms | 60.8 ms |
| Total Pipeline Time* | 387.7 ms | 466.6 ms | 525.3 ms | 625.2 ms |
The measurements represent host-side video-processing performance rather than AI inference performance. Actual inference performance varies depending on the selected accelerator architecture, AI model, model precision, SDK and runtime version, and workload configuration.
Purpose-Built for Multi-Stream Edge Vision Applications
The AIE-KR1B-B1 is designed for edge AI workloads in which multiple video sources must be processed locally before or alongside AI inference. Target applications include video surveillance, manufacturing inspection, retail analytics, traffic monitoring, logistics safety, smart infrastructure and edge AI gateways.
In one real-world deployment, a major highway operator deployed the AIE-KR1B-B1 with an Axelera AI Embedded 110m accelerator for a parking management application , replacing failure-prone legacy sensors with camera-based vehicle detection and classification. Vehicle recognition accuracy exceeded 95 per cent post-deployment, surpassing the customer’s original desktop GPU setup, while cutting operating and maintenance costs.
“Edge AI workloads continue to grow while power and thermal budgets at the edge remain constrained,” said Chester Chung, Product Account Manager at Aetina. “The AIE-KR1B-B1 combines RK3588-based video processing with expandable M.2 AI acceleration. The host handles video decodes and image processing, while the selected accelerator handles AI inference. This architecture gives customers the flexibility to match accelerator choice to model, performance, power and cost requirements as workloads evolve.”
For more information about the AIE-KR1B-B1 and Aetina’s edge AI Vision Solution, please visit Aetina AI Vision Solution.
For more electronics updates please visit: https://efemag.co.uk/category/news
