How do we convert and optimize custom PyTorch or TensorFlow models for the onboard 6.0 TOPS NPU?

Models must be exported to standard ONNX or TensorFlow Lite formats and then compiled into the proprietary .rknn format using Rockchip's RKNN-Toolkit2. The NPU supports INT4, INT8, INT16, and FP16 mixed-precision operations. For optimal throughput on the RK3588 architecture, we recommend executing INT8 asymmetric post-training quantization (PTQ). This layout triggers the NPU's specialized tensor matrix multipliers, maintaining over 98% inference accuracy for vision models like YOLOv8 or mobilenet_v2 compared to unquantized float32 baselines.

What is the physical layer diagnostic procedure for the native CAN Bus, and how do we verify the '60-Ohm Rule'?

Per the ISO 11898 standard, a high-speed CAN network requires two 120Ω termination resistors placed at the extreme electrical ends of the bus loop. To test this on the AE-3588NBT's Phoenix terminal block using a digital multimeter (DMM): Power down the node entirely to prevent active transceiver bias from disturbing the impedance measurement. Set the multimeter to measure resistance (Ohms) and probe directly across the CAN_H and CAN_L terminals. ~60Ω Readings indicate Normal Topology (two 120Ω resistors in parallel). ~120Ω Readings indicate Under-Terminated loop (missing terminal block or trunk cable break). ~40Ω or Lower signals Over-Terminated network, which dampens signal voltage differentials and creates frame dropouts.

What is the glass-to-glass latency for the video processing pipeline using HDMI 2.0 Input and HDMI 2.1 Output?

The hardware ingestion path from the 4K@60Hz HDMI Input to system memory via the V4L2 (Video for Linux Two) subsystem consumes between 30ms and 45ms of delay. Total loop latency is contingent upon your computational load. For example, processing a lightweight object detection network (e.g., YOLOv8 Nano) under the optimized RKNN runtime adds 8ms to 12ms of inference overhead. Hardware display rendering out through the 8K@60Hz HDMI 2.1 port accounts for less than 5ms, yielding a total deterministic pipeline latency of under 65ms for automated optical inspection (AOI) capture loops.

How does the integrated supercapacitor protect file systems during sudden power-off events?

The AE-3588NBT features an onboard hardware monitoring circuit coupled with a heavy-duty supercapacitor matrix. When primary DC input voltage drops below 9V, an interrupt line triggers the onboard hardware supervisor. The system utilizes the supercapacitor's residual power buffer to execute an automatic emergency shutdown routine. This window provides the Linux/Android storage driver sufficient uptime to unmount active flash blocks, flush write caches to the 128GB eMMC, and terminate crucial SQLite or logging loops, preventing raw file system corruption or block-level bricking on unattended field nodes.

What are the precise grounding deployment constraints for the non-isolated RS232 and RS485 blocks?

Because the 4× RS232 and 2× RS485 serial communication paths share a unified ground plane with the main system power common, you must maintain a single-point equipotential ground network across your control cabinet. Bypassing this step can subject the transceivers to severe ground loops if substantial potential shifts occur relative to target PLCs. If routing data lines beyond 10 meters, or parallel to massive inductive loads (such as high-current contactors or variable frequency drives), external inline opto-isolators must be deployed. For environments requiring native 4000V hardware isolation out of the box, select the AE-3588BT.

How does the passive unibody chassis prevent thermal throttling at the upper limit of its +60°C operating window?

The system utilizes a fanless, extruded aluminum unibody enclosure that doubles as the primary heat sink. The Rockchip RK3588 SoC is coupled to the interior chassis surface via high-conductivity thermal silicon interfaces rated at 6.0 W/m·K. At an absolute ambient boundary of +60°C, under continuous 100% NPU and CPU loads, this thermal design limits the internal silicon junction temperature to 81°C, staying comfortably beneath the hardware-enforced 85°C thermal throttling envelope. This design completely eliminates mechanical failure vectors associated with cooling fans in dusty industrial environments.

Is the provided Linux Board Support Package (BSP) distributed with upstream kernel source code access?

Yes, BITECH provides full root development access. The product ships with a production-grade BSP based on the Linux Kernel 5.10 / 6.1 LTS branch. Your engineering team receives complete access to the kernel source code, hardware cross-compilation toolchains, bootloader binaries, and default defconfig parameters. This allows systems engineers to compile custom out-of-tree kernel modules, insert precise packet filtering rules, patch V4L2 device trees for proprietary MIPI CSI camera configurations, or strip unnecessary daemons to optimize boot cycles.

Can the dual Gigabit Ethernet ports handle network traffic segregation for camera streams and automation uplink?

The 2× Gigabit LAN ports are controlled by independent, physical PCIe-bound Ethernet controllers (MAC/PHY). This dual-MAC topography allows system developers to enforce complete network segregation at the OS level. For instance, eth0 can be assigned to a dedicated local subnet for high-bandwidth GigE Vision industrial cameras or localized fieldbus nodes, while eth1 manages outbound traffic routing to factory MES servers, SCADA architecture, or cloud endpoints. This physically eliminates cross-talk bottlenecks and insulates automation nodes from localized data saturation.

Does the native Android 12 OS support complete root modifications for embedded kiosk architectures?

The system image is compiled with unlocked open root privileges. Unlike retail Android setups, our B2B Android build gives hardware developers root access right out of the box. This enables system integrators to disable standard Android system task bars, declare proprietary user applications as the system launcher for bulletproof kiosk lockdown modes, adjust system-level SELinux permissions, and load compiled native C++ binaries into background startup daemons without encountering security verification flags.

How does BITECH guarantee supply chain continuity and revision control for long-term multi-year contracts?

We provide a strict 5 to 7 year lifecycle guarantee for the AE-3588NBT computation block. BITECH operates under rigid product revision control regulations; any unavoidable hardware alterations triggered by component obsolescence (such as LPDDR4 memory layer or eMMC module updates) generate a formal Product Change Notification (PCN) distributed exactly 6 months before execution. This window gives your engineering team ample time to procure final inventory buffers of the previous revision or run hardware-in-the-loop qualification tests on the incoming revision.

Edge AI ComputerAE Platform6 TOPS NPU

AE-3588NBT: RK3588 Fanless Industrial AI Gateway | 6 TOPS NPU

Name: BITECH AE-3588NBT RK3588 Fanless Industrial AI Gateway
Brand: BITECH
SKU: AE-3588NBT
Availability: InStock

A versatile and cost-effective indoor edge AI gateway designed for machine vision, smart infrastructure, and AGV applications in challenging environments.

Verify AI Configuration Download Datasheet

AE-3588NBT-RK3588-Industrial-AI-Gateway-Fanless-Box

Cost-Effective ARM Edge Computing Node for Smart Infrastructure & Kiosks

The AE-3588NBT transforms the powerful RK3588 architecture into a production-ready edge gateway, balancing high-efficiency 6 TOPS AI inference with a rich set of industrial I/O for vision-based automation and smart infrastructure.

Flagship RK3588 Processing

Octa-core processor (4× A76 + 4× A55) up to 2.4GHz delivers workstation-class performance in a compact, low-power edge form factor.

Standardized 6.0 TOPS Onboard NPU Inference Engine

Built-in NPU with INT4/INT8/INT16/FP16 support. Compatible with TensorFlow, PyTorch (ONNX), and Caffe for rapid model deployment.

Dynamic Multi-Stream 8K Decode & 4K Video Input Pipeline

HDMI 2.1 8K@60Hz output and HDMI 2.0 4K@60Hz input enables simultaneous high-res capture and display for vision workflows.

Industrial-Grade Reliability

Fanless, cableless, jumper-less unibody design rated for 24/7 operation with 15G shock and 5-500Hz vibration resistance.

Developer-First Software

Native Android 12 with full root access. BSPs available for Ubuntu 20.04 and Debian 11 with kernel source code.

Rich Fieldbus Integration: 4× RS232, 2× RS485 & Native CAN Bus

Dual Gigabit Ethernet, CAN Bus, 4× RS232, and 2× RS485 for legacy PLC bridging and Industry 4.0 gateway deployments.

AE-3588NBT Real-World Applications

Purpose-built for vision-based automation and smart edge deployments where AI inference and industrial connectivity are essential.

Industrial Machine Vision & AOI

Real-time defect detection and quality assurance using the HDMI input for 4K camera streams and the NPU for automated optical inspection.

Smart Retail & Digital Signage

Driving 8K ultra-HD displays for interactive kiosks while performing edge-based audience analytics and traffic counting.

Traffic & Infrastructure Monitoring

Reliable outdoor operation for vehicle identification and license plate recognition in smart city deployments.

AGV / AMR Navigation

Using the integrated CAN Bus and RS232/485 ports to interface with motor drivers and sensors for autonomous navigation and control.

Standardized Edge AI Inference & Model Quantization

Built-in 6.0 TOPS NPU provides dedicated AI power for 24/7 industrial performance with mainstream framework support.

AI Performance

Built-in NPU

6 TOPS

Supports mixed-precision inference (INT4/INT8/INT16/FP16) for efficient model deployment.

Supported Heterogeneous AI Frameworks & SDK Toolkits

TensorFlowTensorFlow LiteONNXCaffeMXNetDarknet

RKNN-Toolkit2 provides model conversion, quantization, and optimization for rapid deployment of YOLOv8 and other vision models.

Comprehensive 8K/4K Media Pipeline

High-bandwidth visual workflows with HDMI 2.1 8K output and HDMI 2.0 4K input enabling simultaneous high-res capture and display.

Display & Capture Capabilities

HDMI 2.1 output supports 8K@60Hz for high-resolution monitoring and signage
HDMI 2.0 input captures 4K@60Hz video for AI-based visual inspection
48MP MIPI camera interface for high-precision machine vision
8K@60fps H.265/VP9 decode, 8K@30fps H.265/H.264 encode

Vision Use Cases

Real-time AOI with 4K camera input and NPU defect detection
Smart retail kiosks with 8K display and audience analytics
Traffic monitoring with license plate recognition
Digital signage with simultaneous capture and display

Hardened Android Operating System & Fail-Safe Fleet Operations

Designed for 24/7 unattended deployment with comprehensive remote management capabilities.

Open-Root Android 12 & Linux BSP Deep Kernel Access

Native Android 12 with full root access for deep system-level customization and proprietary driver integration.

Automated System Recovery via Hardware & Software Watchdog

Custom bootloader and OTA protocols for remote system/application upgrades across distributed deployments.

Supercapacitor Circuit Design for Robust Power-Off Protection

System Reliability Features

Built-in watchdog (software + hardware)

RTC with backup battery

Supercapacitor for power-off protection

Scheduled power on/off support

100% 18-hour power-on burn-in test under load

AE-3588NBT Technical Specification Blueprint (DC 9-36V Wide-Range Power)

A. Core Processing Engine & Memory Topography

Processor	RK3588 Octa-core (4× A76 @ 2.4GHz + 4× A55 @ 1.8GHz)
GPU	Mali-G610 MP4 (OpenGL ES 3.2, 450 GFLOPS)
NPU	6.0 TOPS (INT4/INT8/INT16/FP16 precision)
Memory	4GB LPDDR4 (8GB/16GB/32GB options)
Storage	128GB eMMC (256GB option)

B.Embedded OS Customization & Unattended Operational Stability

Operating System	Android 12, Debian 11, Ubuntu 20.04
Root Access	Native Android OS with Open Root Access
Stability Features	Hardware/Software Watchdog, RTC with backup battery
Development Support	Full BSP, kernel source, RKNN-Toolkit2

C. High-Bandwidth Video Pipelines & Physical Interface Layout

Display Output	1× HDMI 2.1 (8K@60Hz)
Display Input	1× HDMI 2.0 (4K@60Hz capture)
Camera Interface	1× MIPI (48MP support)
Serial Ports	1× CAN Bus, 4× RS232, 2× RS485/232
GPIO	4-Bit Digital I/O
USB	2× USB 3.2 Gen1, 2× USB 2.0
Network	2× Gigabit LAN (1000Mbps)
Wireless	WiFi 6 (optional), BT 5.2, 4G/5G via Mini PCIe

D.Wide-Range Power Management & Environmental Protection Standards

Power Input	DC 9-36V Wide Range (Phoenix Terminal)
Operating Temperature	-10°C to +60°C
Storage Temperature	-30°C to +70°C
Shock Resistance	15G / 11ms
Vibration Resistance	5–500 Hz / 1.5 Grms
Certification	CE, FCC Class A, RoHS

E. Fanless Unibody Mechanical Dimensions & Mounting Options

Dimensions	190 × 126 × 32 mm
Weight	1.3 kg
Mounting	VESA 75/100, DIN Rail, Wall Mount
Enclosure	Fanless aluminum unibody, IP40

This overview highlights decision-level specifications. For full electrical and mechanical details, refer to the datasheet.

BITECH Industrial ARM Platform Matrix: AE-3588NBT vs. BT vs. LBT

Criterion	AE-3588NBT	AE-3588BT	AE-3588LBT
AI Performance	6 TOPS (onboard only)	Up to 58 TOPS (6 + 52 expansion)	Up to 32 TOPS (6 + 26 expansion)
Network	2× GbE LAN	3× GbE LAN + 5G/4G	2× GbE LAN
Industrial Interface Isolation	Non-isolated	4000V optical isolation on all I/O	Non-isolated
Operating Temperature	-10°C to +60°C	-40°C to +80°C	-40°C to +80°C
Best For	Smart retail, indoor vision, AGV	ESS, substations, harsh environments	Industrial control automation cabinets and equipment with built-in AOI vision inspection units

Engineering Note

AE-3588NBT is ideal for cost-effective smart infrastructure, indoor vision, and AGV applications. For extreme temperatures or high EMI/ground loop risk, consider AE-3588BT with 4000V isolation.

OEM & System Customization Services for Device Integrators

Designed for system integrators and OEMs who require platform customization and long-term supply chain stability for smart infrastructure projects.

OS customization (Android / Linux / Ubuntu / Debian)

Fieldbus and I/O configuration customization

AI framework and deployment support

Suitable for series production and OEM integration

BOM freeze and controlled revision management

Standardized B2B Project Engineering Timeline

Technical Requirement Alignment & Scope Definition (1-2 Weeks)

1-2 weeks

I/O, OS, and integration scope definition

Hardware-in-the-Loop Evaluation & Sample Validation (2-4 Weeks)

2-4 weeks

Hardware validation and software integration

Custom BSP Cross-Compilation & Out-of-Tree Driver Implementation (4-8 Weeks)

4-8 weeks

BSP, drivers, and system configuration

Quality Assurance Testing & Volume Production Ramp-Up (2-4 Weeks)

2-4 weeks

Quality validation and volume production

Why Choose BITECH for Industrial Integration?

Reliable RK3588 Implementation with Production-Ready Stability for Professional Scale Industrial Deployment

Choosing BITECH AE-3588NBT provides your project with a stable, production-ready RK3588 platform that eliminates the complexities of consumer-grade boards, ensuring your AI vision algorithms have the 6 TOPS of dedicated power needed for 24/7 industrial performance.

Optimized CAN Bus RS232 RS485 Connectivity for Legacy PLC and Modern AI-Driven Cloud Gateway Bridging

By integrating CAN Bus, RS232, and RS485 into a single fanless chassis, BITECH enables engineers to bridge the gap between traditional PLC systems and modern AI-driven cloud gateways, streamlining the deployment of Industry 4.0 solutions.

Fanless Unibody Cableless Design with Supercapacitor Power Protection for Unattended Remote Stability

The AE-3588NBT features a unibody, cableless physical design and an integrated supercapacitor for improved power-off protection, ensuring that remote installations in smart cities or factories remain operational without frequent manual intervention.

5-7 Year Long-Term Supply Commitment with Expert CAN RS485 SDK Customization for Accelerated Time-to-Market

BITECH secures your investment with a guaranteed 5-7year product lifecycle and a 24-month warranty, backed by expert technical support for CAN/RS485 debugging and SDK customization to accelerate your time-to-market.

Adjacent Models for Platform Comparison

Compare adjacent models to validate your selection.

AE-3588BT

Best for: ESS sites and harsh field environments

4000V isolated I/O, up to 58 TOPS

View Product

AE-3588LBT

Best for: Cabinet-level AI with fieldbus integration

Non-isolated I/O for controlled environments

View Product

Technical FAQ

AE-3588NBT engineering questions

How do I pick between the AE-3588NBT, LBT, and BT?

All three share the RK3588 core but target different environments. The AE-3588NBT is the cost-effective indoor gateway: 6 TOPS fixed NPU, non-isolated I/O, -10°C to +60°C — ideal for smart retail, signage, indoor vision, and AGV. The AE-3588LBT adds M.2 NPU scaling to 32 TOPS and wide-temp for cabinet AOI. The AE-3588BT adds 4000V isolation, triple GbE and 58 TOPS for substations and ESS. Send us your environment and we'll point you to the right one.

The NBT has no isolation — can I add it, or do I need a different model?

The AE-3588NBT's RS232/RS485/CAN share a common ground, so for clean indoor cabinets with a unified ground plane it's the right, lower-cost choice. For long cable runs (over 10 m), high-EMI sites, or field devices without a shared ground, you either add external inline opto-isolators, or — better — choose the AE-3588BT with native 4000V isolation. We don't put isolation on this model because its target deployments don't need it — that's how it stays cost-effective.

How do I convert PyTorch or TensorFlow models for the 6 TOPS NPU?

Export your model to ONNX or TensorFlow Lite, then compile to the .rknn format with Rockchip's RKNN-Toolkit2 (PyTorch goes through ONNX export — it isn't loaded natively). The NPU supports INT4/INT8/INT16/FP16; for best throughput on RK3588 we recommend INT8 asymmetric post-training quantization, which retains over 98% accuracy on vision models like YOLOv8 and MobileNetV2 versus FP32 baselines.

How do I verify CAN bus termination (the "60Ω rule")?

Per ISO 11898, terminate the bus with 120Ω at both extreme ends. With the node powered down, measure resistance across CAN_H and CAN_L at the Phoenix terminal: ~60Ω = two terminators correctly in parallel (compliant to 1 Mbps); ~120Ω = under-terminated, a missing remote terminator or cable break; ~40Ω or lower = over-terminated, extra resistors clamping the differential and dropping frames.

What's the glass-to-glass latency for the 4K-in / 8K-out pipeline?

HDMI 2.0 4K@60Hz input to system memory via V4L2 runs 30–45 ms; a YOLOv8 Nano under RKNN adds 8–12 ms; HDMI 2.1 8K output rendering is under 5 ms — total deterministic pipeline under 65 ms, within AOI capture-loop windows.

How does the supercapacitor protect the eMMC on sudden power loss?

When DC input drops below 9V, a hardware supervisor interrupts the SoC and the onboard supercapacitor buffers enough power for a clean shutdown — flushing write caches and unmounting flash blocks on the 128 GB eMMC before power is gone. This prevents file-system corruption on unattended retail or field nodes.

How does the fanless chassis avoid throttling at +60°C?

The extruded aluminum unibody acts as the heatsink, coupled to the RK3588 via 6.0 W/m·K thermal interfaces. At +60°C ambient under full CPU+NPU load, junction temperature stays near 81°C — below the 85°C throttle threshold — with no fan to fail in dusty environments. Note this model is rated -10°C to +60°C; for -40°C to +80°C wide-temp, see the LBT or BT.

Is the Linux BSP shipped with kernel source for custom modules?

Yes. The AE-3588NBT ships an open-root BSP on Linux Kernel 5.10 / 6.1 LTS with full kernel source, cross-compilation toolchains, bootloader binaries, and defconfigs. You can compile out-of-tree modules, patch V4L2 device trees for proprietary MIPI cameras, or strip daemons to optimize boot. Native Android 12 also ships with unlocked root for kiosk-lockdown launchers and SELinux adjustments.

How long is lifecycle support and how are changes handled?

5-7 years lifecycle, Any unavoidable component change (e.g. LPDDR4 or eMMC revision) triggers a formal PCN issued 6 months ahead, giving you time for last-time buys or re-qualification. the 18-hour full-load burn-in applies to this model.

Evaluate the AE-3588NBT

The cost-effective RK3588 gateway — matched to your indoor deployment.

Smart retail, signage, indoor vision, or AGV in a clean cabinet? The AE-3588NBT gives you 6 TOPS, an 8K/4K media pipeline, and full serial + CAN fieldbus at the lowest cost in the RK3588 family. Send us your displays, cameras, and I/O, and an engineer returns a configuration and sample terms.

8K HDMI-out + 4K-in pipeline validated for your signage or AOI workflow
Fieldbus plan: native CAN + 4× RS232 + 2× RS485 for PLC bridging
Open-root Android/Linux BSP and sample units for evaluation

An engineer — not a sales bot — replies within 24 hours

Talk to engineering

Configuration review, media/fieldbus sizing, and sample requests for the AE-3588NBT.

Discuss your configuration Request the datasheet

MOQ & volume pricing on request
Sample units available for evaluation

Need wide temp, more TOPS, or isolation? Step up to the AE-3588LBT (32 TOPS, -40°C to +80°C) or the AE-3588BT (58 TOPS, 4000V isolated).

Not sure which RK3588 read the selection guide