Edge AI Box · Intel Core i9-13900 · RTX MXM GPU

AE-760EBT: Intel® Core™ i9 + NVIDIA® RTX MXM Industrial Edge AI PC

13th Gen Intel® Core™ i9-13900 (24C/32T) + NVIDIA RTX 4060/4070/4090/5080/5090 MXM, actively cooled by dual 80 mm industrial fans — desktop-class AI inference for machine vision, AMR and edge generative AI.

  • NVIDIA RTX MXM discrete GPU (up to RTX 5090) for CUDA / TensorRT inference
  • 4× 2.5GbE Intel i226-IT, 2× CAN FD, 8 DI / 8 DO isolated for vision & automation
  • 9–36V wide voltage + remote power, dual 80 mm PWM fans, -10°C to +55°C 24/7 operation

AE-760EBT RTX MXM Industrial PC

AE-760EBT  MXM Industrial AI BOX PC Overview

i9-13900 + RTX MXM, Engineered with Industrial Active Cooling

The AE-760EBT bridges the long-standing gap between consumer-grade GPU workstations and rugged industrial controllers. It combines the 13th Gen Intel® Core™ i9-13900 processor (24C/32T, 8 Performance-cores + 16 Efficient-cores, up to 5.6GHz, 65W base / 219W turbo power) with NVIDIA RTX MXM 3.1 modules ranging from RTX 4060 to RTX 5090 on a single industrial-grade motherboard.

To sustain full P-core turbo performance of the i9 processor together with continuous MXM GPU workloads under intensive AI computing, the chassis is equipped with dual 80mm industrial-grade ball-bearing PWM fans dedicated to a copper heat-pipe thermal array. The cooling system is engineered specifically for 24/7 continuous operation, featuring fan-failure monitoring and tachometer feedback, rather than conventional consumer-grade cooling solutions.

For machine vision builders, the four 2.5GbE i226-IT ports aggregate multiple GigE Vision cameras directly into a single inference node. For AMR / AGV and vehicle integrators, 9–36V wide-voltage input, dual CAN FD, isolated DIO and a remote power terminal allow drop-in integration with existing wiring harnesses. For outdoor and roadside edge deployments, the -10°C to +55°C operating range, IEC 61000-4 immunity and replaceable industrial fans enable serviceable, high- throughput deployment in ventilated enclosures.

The AE-760EBT is a single platform that replaces three traditional building blocks: an industrial PC, a GPU workstation, and an external I/O / network gateway.

AE-760EBT Key Engineering Highlights


Scalable AI Inference: NVIDIA® RTX™ MXM Discrete GPU Architecture

RTX 4060 / 4070 / 4090 / 5080 / 5090 MXM modules — up to 80+ TFLOPS for CUDA, TensorRT and Tensor Core AI inference.


Desktop-Class Processing: 13th Gen Intel® Core™ i9-13900 (24C/32T)

24C/32T (8 P + 16 E), up to 5.6 GHz, 36 MB L3, DDR5 up to 64 GB, TPM 2.0 — desktop-class CPU for AI pre/post-processing.


High-Throughput Vision Networking: Quad 2.5GbE Intel® i226-IT Ports

Quad Intel i226-IT controllers for Muti-camera aggregation, network segmentation and PTP-grade determinism.


Field-Bus Mastery: Native Dual CAN FD & Isolated Digital I/O

2× CAN FD, 4× isolated RS232/485, 8 DI / 8 DO opto-isolated (3.3V/5V/12V) for direct field-bus integration.


Rugged Industrial Chassis

Extruded aluminum frame with copper heat-pipe arrays from CPU and MXM GPU to a finned heat exchanger — built for serviceability.


9–36V + Remote Power

DC jack + 4-pin Phoenix + 2-pin remote power on/off — direct fit for 12V/24V vehicle and cabinet wiring.


Active Thermal Engineering: Dual 80mm Industrial PWM Fans

2× 80 mm (8 cm) ball-bearing PWM fans, >70,000 hrs MTBF, fan-fault & tachometer monitoring — sustains full i9 + RTX load.

AE-760EBT Technical Specifications (Key Components & Interfaces)

ProcessorIntel® Core™ i9-13900 (24C/32T: 8 P-cores + 16 E-cores, up to 5.6 GHz, 36MB L3, 65W base / 219W turbo)
Integrated GPUIntel® Iris® Xe / UHD Graphics
Discrete GPU (MXM)NVIDIA RTX 4060 / 4070 / 4090 / 5080 / 5090 MXM (optional)
CoolingActive cooling – 2× 80 mm (8 cm) industrial ball-bearing PWM fans, >70,000 hrs MTBF, fan-fault & tachometer monitoring
Memory2× DDR5 SO-DIMM, up to 64 GB
Storage1× 2.5″ SATA SSD/HDD + 1× mSATA
Display2× HDMI 2.0 Type-A, up to 4096 × 2160 @ 30 Hz
Ethernet4× Intel i226-IT 2.5GbE LAN
Serial4× RS232/485 (Isolated)
CAN Bus2× CAN FD
Digital I/O8× DI + 8× DO, opto-isolated (3.3V / 5V / 12V)
USB4× USB 3.2 Gen1 Type-A (5 Gbps)
Wireless / WWAN1× M.2 B-Key (4G/5G + SIM), 1× M.2 E-Key (Wi-Fi/BT)
SecurityTPM 2.0
Power InputDC 9–36V (jack 5.5×2.5 mm + 3-pin Phoenix), Remote Power On/Off
Operating Temp.-10°C to +55°C (PassMark BurnInTest at full i9 + RTX load, no thermal throttling)
Shock / Vibration15G @ 11ms / 5–500 Hz, 1 Grms
Dimensions / WeightCompact active-cooled chassis with dual 80 mm fan exhaust / approx. 6.0 kg
MountingWall / Side / DIN-rail / VESA
OS SupportWindows 11 Pro, Ubuntu, Linux

This overview supports early-stage system verification. For full electrical, mechanical, EMI/EMS and environmental specifications, please refer to the datasheet.

i9-13900 + RTX 4060 MXM — Heavy-Load Inference Compute & Power-Efficiency

Measured engineering reference for the AE-760EBT default SKU (Intel® Core™ i9-13900 + NVIDIA® RTX 4060 MXM, 8 GB GDDR6). Values reflect sustained heavy-load inference (CUDA + Tensor Core, batch ≥8) inside the industrial-grade chassis with dual 80 mm PWM cooling at 25 °C ambient. Use these as a planning baseline; final numbers depend on model, batch size and thermal envelope.

Metrici9-13900 (CPU)RTX 4060 MXM (GPU)System (CPU + GPU)
Peak compute~0.9 TFLOPS FP32 (AVX2/AVX-512 disabled, 24C/32T)15.1 TFLOPS FP32 / 121 TFLOPS FP16 Tensor / 242 TOPS INT8≈ 242 TOPS INT8 effective for vision inference
Sustained heavy-load power125–180 W (PL1 125 W, PL2 219 W, sustained ~150 W typical)95–115 W (MXM TGP capped at 115 W in chassis)260–320 W under joint CPU+GPU inference load
Idle / light-load power18–25 W package8–12 W (P8 idle)35–55 W full system idle
Inference power-efficiency~6 GFLOPS/W (FP32, AVX2)≈ 2.1 TOPS/W INT8 sustained≈ 0.85 TOPS/W INT8 measured at the wall
YOLOv8-s @ 640×640 (TensorRT INT8)~38 FPS (OpenVINO, single stream)~410 FPS (batch=8, INT8)≈ 1.4 FPS / W (camera-stream throughput)
ResNet-50 inference (ImageNet, INT8)~310 img/s~3,950 img/s (batch=32, TensorRT)≈ 13.5 img/s / W
LLaMA-2 7B (INT4, llama.cpp / TensorRT-LLM)~6 tokens/s (CPU only, Q4_K_M)~58 tokens/s (8 GB VRAM, batch=1)≈ 0.20 tokens/s / W end-to-end
Stable Diffusion 1.5 (512×512, 20 steps)Not practical (>40 s/image)~2.4 s/image (FP16, TensorRT)≈ 1 image per 720 J of wall-plug energy
Thermal headroom (dual 8 cm PWM, 25 °C amb.)Tj ≤ 92 °C @ 180 W sustainedTj ≤ 78 °C @ 115 W sustainedNo thermal throttling observed up to 320 W system load
Acoustic level @ 1 m (full inference load)≈ 48 dB(A) — both 80 mm fans at ~70% PWM
Best perf / W
≈ 2.1 TOPS / W
RTX 4060 MXM, INT8 sustained inference at 110 W TGP.
Typical heavy-load draw
260 – 320 W
Joint CPU+GPU inference; size 9–36 V DC supply for ≥ 400 W margin.
Vision throughput
~410 FPS YOLOv8-s
INT8, batch=8, 640×640 — supports 8+ concurrent 2.5GbE camera streams.

Figures derived from Bitech engineering bench data (TensorRT 9.x, CUDA 12.x, Ubuntu 22.04, kernel 6.5) on the AE-760EBT reference unit. Customer workloads, model precision and

ambient temperature will shift absolute numbers — contact our engineering team for an application-specific power & thermal budget review.

  AE-760EBT Target B2B Application Scenarios

Multi-Camera Machine Vision & Automated Optical Inspection (AOI)

GigE Vision aggregation across 4× 2.5GbE LAN with RTX-accelerated YOLO / defect-detection inference for SMT, semiconductor and packaging lines.

Onboard AI Controller for AMR / AGV & Autonomous Vehicles

9–36V vehicle power, CAN FD motion bus and 3D point-cloud / SLAM acceleration with i9-13900 + RTX MXM on a single rugged active-cooled unit.

Edge Generative AI & Infrastructure Operations

Run Stable Diffusion, LLaMA-class and ONNX models locally with TensorRT on RTX 4090 / 5090 MXM — no cloud dependency.

Roadside & Outdoor V2X / ITS

Wide-temperature, wide-voltage, IEC 61000-4 immunity for traffic vision, ANPR and intersection edge AI.

Digital Twin & 3D Visualization HMI

Dual 4K HDMI outputs driven by discrete RTX GPU for CAD, BIM and operator-facing twins on the plant floor.

Robotics & Service Robot Compute

Compact actively-cooled chassis, isolated DIO and CAN FD for humanoid, inspection and service robot platforms requiring i9-class compute.

Technical Evaluation: AE-760EBT (x86) vs. AE-NJ60BT (ARM Jetson)

Both are active-cooled edge AI platforms, but they target different compute classes and software ecosystems. Use the AE-760EBT when you need x86 + CUDA-class discrete GPU; choose the AE-NJ60BT when you need an ARM Jetson Orin SoC with native GMSL2 camera ingress at lower power.

Engineering FactorAE-760EBTAE-NJ60BT
Architecturex86 — Intel Core i9-13900 (24C/32T)ARM — NVIDIA Jetson Orin Nano / NX / NX Super SoM
AI ComputeNVIDIA RTX 4060–5090 MXM, up to 80+ TFLOPS / Tensor & RT coresUp to 157 TOPS (Orin NX Super 16GB), integrated GPU + DLA
Software StackWindows 11 Pro / Ubuntu, CUDA, TensorRT, OpenVINO, ONNXNVIDIA JetPack, CUDA, TensorRT, DeepStream
Camera Ingress4× 2.5GbE GigE Vision + USB 3.24-lane MIPI CSI or 4× GMSL2 (long-distance, EMI-resistant)
CoolingActive — dual 80 mm industrial PWM fansActive — 1× smart PWM fan (low-noise, temperature-controlled)
Operating Temp.-10°C to +55°C-40°C to +70°C (wide-temp, single-fan assisted)
Power InputDC 9–36V, remote on/offDC 9–36V, remote on/off
Best ForHeavy multi-stream vision AI, generative AI on the edge, CUDA workstations in the fieldCompact multi-camera AMR/AGV perception, autonomous machines, in-cabinet vision AI
Running a large language model at the edge is bound by two things: how much memory holds the model, and how fast the accelerator generates tokens. That splits neatly into three tiers of BITECH hardware.

(Figures are 4-bit (Q4/INT4) single-stream generation; actual throughput varies with quantization, context length and runtime. Sources: NVIDIA JetPack 6.2 LLM benchmarks and independent edge-LLM inference testing.)

PlatformRuns well (4-bit)Practical ceiling7–8B speed (Q4)Memory for the modelBITECH box
x86 + RTX MXM GPU7B–34B70B quantized~100+ tok/s16–32 GB discrete VRAMAE-760EBT
Jetson Orin (CUDA GPU)2B–8B13B (AGX 64 GB)~15–40 tok/sup to 64 GB unified LPDDRAE-NJ60BT
RK3588 (6 TOPS NPU)1B–4B~7B (slow)~3–15 tok/sup to 32 GB shared LPDDRAE-3588BT

Designed for 24/7 Industrial & Vehicle Deployment

  • Operating temperature: -10°C to +55°C with full i9-13900 turbo + RTX load (storage -20°C to +60°C)
  • Active cooling: 2× 80 mm (8 cm) ball-bearing PWM fans, >70,000 hrs MTBF, fan-fault & tach monitoring, field-replaceable
  • Shock 15G @ 11ms half-sine; vibration 5–500 Hz, 1 Grms
  • EN/IEC 61000-4-2/3/4/5/6/8/11 immunity, CISPR 32 / EN 55032 Class A
  • 100% 18-hour PassMark BurnInTest under full CPU + GPU load
  • 10 year lifecycle, 24-month warranty, PCN/EOL management

Specifications are based on factory validation. Final environmental performance should be confirmed against the customer’s installation profile.

Related Products

Compare adjacent models to validate your selection.

AX-660EBT

Best for: PCIe x4 modular expansion

Internal PCIe slot for serial/LAN/AI cards without discrete GPU

View Product

AX-530EBT

Best for: Legacy PCI/PCIe expansion

Multiple PCI + PCIe slots for legacy motion/DAQ cards

View Product

AE-NJ60BT

Best for: ARM-based 157 TOPS edge AI

NVIDIA Jetson Orin NX Super with native GMSL2/MIPI camera ingress

View Product

Technical FAQ

How does the AE-760EBT handle the Core i9-13900 power limitations and thermal throttling under full AI inference load?

The motherboard BIOS applies optimized power profiles tailored specifically for industrial deployments: Long-term Power Limit (PL1) is locked at 125W, while the Short-term Power Limit (PL2) allows bursts up to 219W for a duration of 28 seconds.

To sustain this without thermal dropping, the system utilizes 4x composite copper heat pipes mated directly to the CPU core, drawing heat to an aluminum fin array cooled by twin 80mm high-static-pressure PWM fans. In a 55°C ambient environment running continuous validation workloads, the CPU core temperature stabilizes below 92°C, allowing the P-cores to maintain a steady 3.0GHz to 3.5GHz baseline clock indefinitely.

Are the four 2.5GbE network ports on the AE-760EBT independent or bridged? Will they fight for bandwidth under full line rate?

They are completely independent. We do not use low-cost single-channel switch chips to bridge network traffic. Instead, the mainboard integrates 4x independent Intel® i226-IT industrial-grade 2.5GbE controllers, each occupying a dedicated PCIe lane.

When aggregating four high-bandwidth GigE Vision cameras running simultaneously at full 2.5Gbps line speed, the total system throughput approaches 10Gbps. This native architecture prevents internal bus congestion, packet collision, and dropped frames at the edge.

How are the 4-channel GMSL2 camera interfaces supported on the AE-NJ60BT BSP?

The standard out-of-the-box Board Support Package (BSP) pre-integrates the device tree overlay for the onboard MAX9296 deserializer chipset.

  • Native Support: For industry-standard vehicle and industrial vision sensors (e.g., IMX390, AR0231), we provide compiled .dtb files and kernel patches file directly.
  • Third-Party Integration: If you are deploying proprietary sensor modules, BITECH supplies the complete BSP source repository. Alternatively, our hardware engineering lab can develop customized sensor drivers for your deployment.
Our AMRs experience severe voltage drops during heavy motor acceleration. What hardware protections are built into the DC input stage?

Both edge deployment platforms feature an isolated wide-range DC 9V to 36V power stage equipped with reverse-polarity protection, over-current cutoff, and transient voltage suppression (TVS) diodes.

To handle typical 12V or 24V lead-acid/lithium battery voltage drops caused by high-torque motor startup (where bus voltage dips down to 9V for up to 100ms), the onboard DC-to-DC converters and high-capacitance solid-state capacitor banks sustain smooth internal voltage output, preventing sudden system resets or power-off failures.

The AE-760EBT lists support up to an NVIDIA RTX 5090 MXM card. What is the maximum sustained TGP the chassis thermal design can handle?

Due to the physical volume constraints of a compact, active-cooled industrial chassis, the mainboard hard-caps the power allocation for the MXM 3.1 Type-B slot at a maximum Total Graphics Power (TGP) of 115W.

When selecting an RTX 4060 MXM (115W), it runs unthrottled at maximum factory clock rates. If configuring for flagship RTX 5080 or RTX 5090 MXM modules, the GPU must be throttled to the 115W power tier via custom VBIOS or driver constraints. This preserves full Tensor Core INT8 inference logic metrics while keeping system thermal profiles stable.

Are the twin CAN FD ports optically isolated? How does the mainboard handle signal reflections on long cable runs?

Yes, they feature 4000V magnetic and signal isolation arrays to isolate ground loops and high-voltage spikes caused by heavy industrial equipment or switching contactors.

For impedance balancing across extended field-bus line drops, the mainboard features onboard 120Ω termination resistors accessible via physical jumper pins. This complies with standard common-mode rules (60Ω loop law) and eliminates reflection-induced data loss without requiring external line terminators.

For unventilated outdoor control cabinets, should we specify the fanless ARM-based AE-NJ60BT or the active-cooled x86 AE-760EBT?

Selection depends entirely on your software framework and enclosure type:

  • Specify AE-NJ60BT (ARM): Best for completely sealed, dust-prone enclosures. It operates entirely on passive chassis conduction across a wide -40°C to +70°C thermal layer, provided your pipeline is compiled natively for JetPack/TensorRT.
  • Specify AE-760EBT (x86): Required if running traditional Windows 11 workloads, complex C++ logic frameworks, or raw PyTorch pipelines. Because it utilizes dual 80mm exhaust fans, the deployment cabinet must feature filtered air intake paths or louvers to maintain airflow over the outer fin blocks.
Is the TPM 2.0 module on the AE-760EBT hardware-based, and do you support factory-level secure boot custom key flashing?

It is a discrete, hardware-level security chip utilizing the Infineon SLB9670 platform, bypassing the security vulnerabilities associated with CPU-simulated firmware TPM (fTPM).

For high-security industrial grid or critical national infrastructure applications, BITECH supports factory-level UEFI Secure Boot customization. We can securely flash your proprietary encryption signatures (PK, KEK, db keys) directly into the custom BIOS firmware matrix during production, locking the operating state against physical firmware injection attacks.

What is the MTBF of the dual 80mm cooling fans, and can they be replaced in the field without breaking thermal blocks or voiding the warranty?

The dual 80mm active cooling array employs industrial dual-ball-bearing fan modules with a verified MTBF greater than 70,000 operational hours. They feature native tachometer monitoring th5-7at feeds direct fan-fault alerts straight to system logs or external GPIO pins.

The thermal design is modular: The fan brackets are external to the sealed internal chassis. You can remove four outer plate screws to swap out a fan module in under two minutes without disturbing the primary copper heat pipes, thermal pads, or factory sealing layers.

How does BITECH mitigate supply chain EOL risks and manage BOM changes for long-term multi-year deployments?

BITECH implements a strict BOM (Bill of Materials) Freeze Policy once an OEM/ODM configuration is approved for series integration. All onboard IC components, network controllers, and chassis specifications are selected exclusively from the Intel Embedded Roadmap and industrial component channels ensuring a 10 year lifecycle guarantee.

In the event of an unavoidable sub-tier component termination, we enforce a mandatory 6 to 12-month Product Change Notice (PCN) or End-of-Life (EOL) alert track, allowing our B2B volume accounts ample validation window to secure last-time-buy allocations or verify dropped-in replacement modules.

Can an edge AI box PC run a local LLM on-device?

Yes — an edge AI box PC can run a local LLM fully on-device, and the right platform depends on model size. An x86 box with a discrete RTX MXM GPU (16–32 GB VRAM) runs 7B–34B models at interactive speed and 70B quantized; a Jetson Orin module runs 7–8B models fanless; and an RK3588 board handles small 1–4B assistants. All keep data on-premise with no cloud calls.

Specify the Right MXM GPU for Your Edge AI Project

Discuss RTX 4060 / 4070 / 4090 / 5080 / 5090 MXM options, camera ingress topology and OEM customization with our engineering team.

Scroll to Top
POPUP Form

Contact Us