What is an edge AI box? An edge AI box is a compact industrial computer with a built-in AI accelerator that runs video analytics on site, next to the cameras. It reads streams from existing IP cameras or NVRs, detects events such as fire, smoke, intrusion or missing PPE, and forwards only alerts, not video.
This guide is written by the engineering team of Beijing Huamei Xuntong Technology Co., Ltd. (TouchVision), a manufacturer founded in 2012 that designs and builds industrial displays, embedded boards, HMI panels, industrial gateways and edge AI computers.
Key takeaways
- An edge AI box runs AI video analytics on site and forwards only events, not video.
- “Edge AI box” and “edge AI computer” describe the same class of device; box usually means a compact video analytics unit.
- Plan capacity on concurrent analytics streams (16 HD on BM1684 models), not only on decode channels (32× 1080p).
- TOPS is a peak INT8 figure: useful for comparison, but decode, memory and models also set real capacity.
- Edge beats cloud when privacy, bandwidth or offline reliability matter; cloud suits small, well-connected deployments.
- Existing RTSP cameras and NVRs can usually be reused, and custom models can be deployed from common frameworks.
What is an edge AI box, and how is it different from an edge AI computer?
An edge AI box is a small industrial computer with a dedicated AI processor that performs neural-network inference on site instead of in a cloud data centre; “edge AI computer” describes the same kind of device and is often used for larger or more general-purpose units.
Inside a typical edge AI box you will find an AI system-on-chip that combines CPU cores, a neural processing unit and hardware video decoders, plus memory, storage, network ports and industrial I/O. All current TouchVision edge AI computers and edge AI boxes use SOPHGO processors: the BM1684 (8× Cortex-A53 at 2.3 GHz, 17.6 TOPS INT8) and the BM1684X (32 TOPS INT8 with FP32/FP16/BF16/INT8 mixed precision). TouchVision uses “edge AI box” for compact video analytics units such as the YZ-B011, YZ-B013 and YZ-B016, and “AI edge server” for higher-capacity models such as the YZ-B012 and YZ-B015.
How does an edge AI box perform AI video analytics?
The box pulls video streams from cameras over the local network, decodes them in hardware, runs detection models on each frame, and turns detections into events that are sent to a terminal, VMS or platform and can drive sirens or barriers directly.
- Connect the box to existing IP cameras or NVRs over the LAN. Streams are typically delivered over RTSP, the IETF real-time streaming protocol.
- Enable algorithms per camera and set zones, schedules and thresholds, for example a line-crossing zone at a fence or a helmet check at a gate.
- Decode and infer. Hardware decoders convert H.264/H.265 streams into frames; the AI accelerator runs the models.
- Raise events. Alerts and snapshots go to a TouchVision security and fire-safety terminal, your VMS or a platform via API.
- Act locally. Opto-isolated digital outputs, RS-485 or isolated CAN let the box trigger sirens, barriers or a PLC without a round trip to a server.
TouchVision offers more than 50 ready-made algorithms, including fire and smoke, intrusion and line crossing, crowd gathering, helmets and work clothing, phone use and smoking, illegal parking, blocked fire lanes and e-bikes entering elevators, with scenario packs for fuel stations, commercial kitchens, residential communities, battery charging sheds, factories and warehouses. Custom models can be trained on request.
How many camera channels can one edge AI box handle?
Check two numbers separately: how many streams the box can decode and how many it can analyse at the same time. A BM1684-based box decodes up to 32 channels of 1080p video and analyses up to 16 HD streams concurrently.
Decode capacity tells you how much video the hardware can unpack; analytics capacity tells you how many of those streams can run AI models in real time. The practical figure depends on the algorithm mix, model size, resolution and how many frames per second each algorithm needs. A slow-changing scene such as an overflowing-bin check needs far fewer frames than intrusion detection at a perimeter. Capacity scales by adding boxes, so a site with 40 cameras can be split across several units placed close to the camera groups.
What do TOPS figures mean for an edge AI computer?
TOPS (tera operations per second) is the accelerator’s peak AI throughput, normally quoted at INT8 precision; it is a good first filter but not a direct measure of how many cameras a box can serve.
- Precision matters. INT8 figures are the headline numbers; FP16 or FP32 throughput is lower. The BM1684 is quoted at 17.6 TOPS INT8 and 2.2 TFLOPS FP32.
- Video decode often limits first. If the decoder cannot keep up, extra TOPS are idle.
- Memory and model fit. Larger models, including language models, need memory as well as TOPS. The BM1684X in the YZ-B033 on-premise LLM appliance delivers 32 TOPS INT8 and runs DeepSeek and Qwen models locally at a typical 17 W.
- Efficiency. TouchVision’s SOPHGO-based designs run at under 1 W per TOPS, with passive-cooling options on selected designs.
Edge AI vs cloud video analytics: which should you choose?
Choose edge AI when video must stay on site, alarms must be fast and reliable even if the internet link drops, or uploading many camera streams is impractical; choose cloud analytics when cameras are few, bandwidth is plentiful and you prefer a subscription service.
| Factor | Edge AI box (on-premise) | Cloud video analytics |
|---|---|---|
| Where video is processed | On site, next to the cameras | In a remote data centre |
| Data leaving the site | Only events and snapshots | Continuous video streams |
| Uplink bandwidth | Low | Grows with every camera |
| Response if the internet fails | Local alarms and outputs keep working | Analytics stop until the link returns |
| Privacy and data-protection review | Simpler: footage stays on your network | Requires review of transfer and storage of footage |
| Cost model | Hardware purchase, scales by adding boxes | Recurring per-camera or per-stream fees |
| Model updates | Deployed to each box, centrally managed where possible | Handled by the provider |
In Europe, keeping footage on site can also simplify compliance work under data-protection rules, although every deployment still needs its own assessment. Read more about on-premise AI video analytics.
Which TouchVision edge AI box fits which project?
All BM1684 models share 17.6 TOPS INT8, 32-channel 1080p decoding and 16-stream analytics; they differ in power input, I/O, storage and size.
| Model | Key difference | Power and temperature |
|---|---|---|
| YZ-B011 | 3× GbE, 2× isolated CAN, RS-232, RS-485, 4× opto-isolated DIO | DC 12–36 V; -20 to +70 °C |
| YZ-B012 | Positioned for camera-dense sites; same industrial platform as YZ-B011 | DC 12–36 V; -20 to +70 °C |
| YZ-B013 | Compact 217 × 155 × 48 mm; modular SATA, LTE, Wi-Fi, HDMI | DC 12 V, typical ≤20 W; -20 to +70 °C |
| YZ-B015 | M.2 NVMe plus optional drive up to 2 TB for evidence clips | DC 12 V; -20 to +60 °C |
| YZ-B016 | Direct 220 V AC input, 2 alarm inputs, 2× RS-485 | 220 V AC; -20 to +60 °C |
| YZ-B033 (BM1684X) | 32 TOPS INT8; local LLMs plus 32-channel video analytics | 17 W typical; -20 to +70 °C |
What should you check before buying an edge AI box?
Confirm the cameras and streams, the algorithms, where alerts must go, the installation environment and how models will be maintained, then size the number of boxes from tested capacity.
- Inventory the cameras: number, resolution, codec (H.264/H.265) and whether RTSP streams are available from cameras or the NVR.
- List the use cases per camera and the response needed: alert only, siren, barrier or PLC action.
- Estimate load: streams to analyse at the same time and the algorithms per stream; plan against concurrent analytics, not decode capacity.
- Define integration: VMS, platform API, on-screen terminal, digital outputs, RS-485 or CAN.
- Check the installation: DC or AC power, cabinet temperature, need for wide-voltage input or fanless operation.
- Decide on storage for evidence clips (for example the NVMe-equipped YZ-B015).
- Plan model lifecycle: ready-made algorithms, custom training, or your own models converted from ONNX, PyTorch or TensorFlow.
- Run a pilot on real cameras and lighting, tune thresholds, then roll out.
Why source edge AI hardware from a manufacturer?
A manufacturer that builds the box, the I/O and the display terminal can adapt the whole system to your site instead of selling a fixed appliance. TouchVision has built industrial hardware since 2012 and extended its range from touch displays and panel PCs to gateways, edge AI computers and on-premise AI appliances. Our engineers customise interfaces, enclosures and mounting, firmware and OS images and branding for OEM projects, and can design a custom carrier board around the BM1684X/BM1684 AI edge computing module. Small-batch pilot orders are supported, and standard products carry a 3-year warranty from delivery. For fully custom builds, visit touchvisiontech.com.
Frequently asked questions
What is the difference between an edge AI box and an edge AI computer?
They describe the same kind of device: an industrial computer with an AI accelerator that analyses camera or sensor data on site. “Edge AI box” is usually used for compact video analytics units; “edge AI computer” or “AI edge server” is often used for higher-capacity or more general-purpose models.
Do I need new cameras to use an edge AI box?
Usually not. TouchVision edge AI boxes take standard RTSP streams from existing IP cameras and NVRs, so the camera network can stay as it is.
How many cameras can one edge AI box analyse?
It depends on the processor and the algorithm mix. TouchVision boxes based on the SOPHGO BM1684 analyse up to 16 HD streams at the same time and decode up to 32 channels of 1080p video; the BM1684X-based YZ-B033 can analyse 32 video streams.
Does video leave the site with edge AI video analytics?
No. Analysis runs on the box inside your network, so footage stays on the local network and only events and alarm snapshots are forwarded to the platform you choose.
Can I run my own AI model on an edge AI computer?
Yes. Models from TensorFlow, PyTorch, ONNX, Caffe, MXNet and PaddlePaddle can be converted with the SOPHGO toolchain and deployed on TouchVision boxes, with Docker and Python support.
Sources and standards referenced
- SOPHGO: SOPHON BM1684X tensor computing processor
- SOPHGO SDK documentation: BM1684 and BM1684X term definitions
- IETF RFC 7826: Real-Time Streaming Protocol Version 2.0
- ONNX: Open Neural Network Exchange
- European Commission: Data protection
Need a recommendation for your project?
Send your drawings, quantities and environment through our RFQ form and a TouchVision engineer will reply with a recommended configuration, quotation and sample plan. You can also call or WhatsApp +86 135 0532 6001.