
SXD-X3 Edge Inference Acceleration Card (9 TOPS)
Specification · SXD-X3 · PCIe card
SXD-X3 Edge Inference Acceleration Card is a 9 TOPS edge AI analytics device, nominally supporting 1 channel of real-time video input (1080P, 1 algorithm per ch…
No. sxd-x3-edge-inference-acceleration-card-9-tops · 2026-11-20
Overview
Product Overview
The SXD-X3 Edge Inference Acceleration Card is a 9 TOPS edge AI analytics device, nominally supporting multi-channel real-time video access (1080P, 1 algorithm per channel, frame sampling every 2 seconds). The actual number of supported channels varies with camera resolution, frame sampling interval, and the number of algorithms per channel, and should not be interpreted as a fixed channel count. For selection guidance, see "Computing Power and Access Specifications" below. PCIe plug-in acceleration card that adds AI computing power to existing servers/industrial PCs.
Computing Power and Access Specifications
The nominal "multi-channel" refers to the number of real-time video access channels, with the baseline specification of 1080P, 25fps, 1 algorithm bound per channel, and one frame sampled every 2 seconds. The actual number of supported channels is not a fixed value and varies with three variables:
| Variable | Change | Impact on Computing Power |
|---|---|---|
| Camera resolution | 2 MP → 4 MP | Decoding and inference load per channel increases by approximately 50%–100%, and the number of supported channels decreases accordingly |
| Frame sampling interval | 2 s → 1 s | Frames processed per second doubles, and the number of supported channels is approximately halved |
| Algorithms per channel | 1 → 2 | Inference count doubles, and the number of supported channels is approximately halved |
Selection recommendation: First calculate based on "resolution × frame sampling frequency × algorithms per channel," then reserve a 30% margin. 2 MP + 2-second frame sampling + single algorithm corresponds to the nominal channel count; when switching to 4 MP or 1-second frame sampling, estimate at 50%–60% of the nominal channel count. 9 TOPS is the INT8 peak computing power; actual available computing power is affected by the algorithm model and concurrency strategy.
Dataflow Architecture and Effective Computing Power
Nominal computing power is only an upper limit. What really determines how many channels can be supported is chip utilization. This device is equipped with a reconfigurable dataflow AI chip that compiles the entire algorithm model into a data pipeline, directly connecting operators into a hardware path. Computation is completed as data flows through, without repeated instruction fetching and movement of intermediate results.
| Comparison Item | Instruction Set Architecture (common GPU/NPU) | Dataflow Architecture (this device) |
|---|---|---|
| Computation method | Executes operators one by one, with intermediate results repeatedly read from and written to memory | Operators are compiled into a pipeline; computation completes as data flows through |
| Chip utilization (public industry comparison) | 10%–30% | 65%–95% |
| Effective computing power converted from 8 TOPS | Approximately 0.8–2.4 TOPS | Approximately 5.2–7.6 TOPS |
| Supported channels at equivalent computing power | Baseline | Approximately 2–3× based on median utilization conversion |
| Algorithm deployment | Depends on framework adaptation and manual tuning | In-house compiler; TensorFlow / PyTorch / ONNX models are automatically compiled and deployed end-to-end |

Effective computing power = nominal peak × chip utilization. Taking 8 TOPS as an example: the instruction set approach has approximately 0.8–2.4 TOPS actually available, while the dataflow approach has approximately 5.2–7.6 TOPS — this is also why devices with the same nominal computing power can differ greatly in the actual number of cameras they can support.
For real-world deployment scenarios, this means three things:
- More cost-effective upgrades to existing systems: One device replaces two to three previous devices, with cabinet space and power consumption reduced accordingly
- Reconfigurable algorithms: Changing algorithms does not require changing hardware; compile and deploy, avoiding "having to replace the box whenever the algorithm is upgraded"
- Computing power can run consistently at full capacity: During concurrent decoding and inference of multiple video channels, the pipeline can still maintain high utilization, avoiding "high peak, low actual measured performance"
Note: The utilization values in the table are publicly available industry architecture comparison ranges. Actual measured values for specific models will vary with operator structure, input resolution, and batching strategy; the nominal channel count on this page is given according to the specification of 1080P, 1 algorithm per channel, and frame sampling every 2 seconds.
Algorithm Capabilities (by Scenario Category)
The algorithms compatible with the device are divided into 6 major categories by scenario, covering 460+ sub-scenarios; which specific algorithms are enabled is configured per project and subject to the delivery list.
| Scenario Category | Representative Algorithms | Typical Locations |
|---|---|---|
| Personnel behavior and on-duty status | Area intrusion, line crossing, personnel lingering, loitering, climbing/overfencing, fall detection, on-duty status monitoring, work procedure deviation, helmet/reflective vest/workwear/mask/glove wearing recognition | Plant perimeter, duty room, work area |
| Safety risks and fire protection | Smoke and fire, open flame, smoke, fire lane occupancy, dangerous area intrusion, passage obstruction, gas leakage, electrostatic discharge | Warehouse, power distribution room, kitchen, hazardous chemical area |
| Vehicles and traffic | License plate recognition, illegal parking, wrong-direction driving, traffic flow statistics, non-motor vehicle intrusion, construction vehicle recognition, height limit | School gate, campus entrance/exit, roads |
| Objects and facilities | Material stacking, pallet recognition, shelf vacancy, equipment status, meter readings, conveyor abnormality | Workshop, warehouse, power distribution room |
| Environment and order | Garbage overflow, standing water, dust, road occupation, disorderly stacking | Plant roads, garbage room, construction area |
| Food and kitchen | Kitchen attire, rodent infestation, sample retention testing, tableware disinfection, cold chain temperature | School/enterprise cafeterias |
Note: Algorithm names are handled according to external communication standards; algorithms involving human portraits must be used within the authorized scope; the table lists representative algorithms, and the complete list is provided with the project.
Algorithm Training and Customization (Optional)

If general algorithms cannot cover on-site scenarios, the "AI Algorithm Training Platform" can be provided to train dedicated models using on-site data. After training is completed, quantization generates SG files that are deployed to the inference acceleration card on this device for execution. The platform features a graphical interface and completes the process in four steps: data annotation → model training → model evaluation → model release.
| Stage | Capability | Description |
|---|---|---|
| Data annotation | Manual annotation + intelligent annotation | Data can be uploaded locally,回流 from the site, or selected from platform datasets; automatically splits sample sets and test sets |
| Model training | Segmentation / detection / classification | Supports retraining existing models and first-time training from the model library; customizable images per batch and network scale (large/medium/small) |
| Model evaluation | Specified model + one-click evaluation | Evaluates performance using an independent test set; supports specified GPU acceleration |
| Model release | Quantization + SG file generation | Converts to dataflow expressions recognizable by the acceleration card and deploys to edge devices for execution |
The training platform is standalone software and requires a supporting GPU server (reference configuration: 2×Xeon Gold 6126 / 64GB / 2×RTX 3080Ti), configured per project. Training results depend on sample quality and scenario coverage; it is recommended to validate with a small batch of data before large-scale collection.
Core Features
- AI computing power: 9 TOPS (subject to actual measurement)
- Model: SXD-X3
- Form factor: PCIe board
- Compatible devices: servers, small hosts, industrial PCs
Alarm Data Flow

Deployment Architecture

Deployment and Access
- PCIe acceleration card form factor; install into an existing server to expand AI inference capability
- 9 TOPS computing power, suitable for small-scale algorithm enhancement on the existing platform side
- Reuses server power and chassis, with no additional standalone devices or cabling
- Shares algorithms and management interfaces with the full-device solution for unified O&M
Selection Recommendations
Suitable for projects that already have servers and only want to add a small number of algorithm channels; if standalone deployment without modifying the server is required, prefer the edge analytics station or all-in-one device.
Related Recommendations
- Solution: Smart Campus Security Monitoring Upgrade Solution
- Solution: Smart Classroom Behavior Analytics Solution
References
The following are regulations and standards related to this document for further verification:
- "Technical Requirements for Information Transmission, Exchange, and Control of Public Security Video Surveillance Networking Systems" (GB/T 28181) — Official Source
- "Data Security Law of the People's Republic of China" — Official Source
Applications




Core Features
9 TOPS(以实测为准)
SXD-X3
PCIe 板卡
服务器、小型主机、工控机
Specifications
| Item | Specification |
|---|---|
| Model | SXD-X3 |
| Form Factor | PCIe card |
| AI Compute | 9 TOPS (subject to actual measurement) |
| Compatible Devices | Servers, small-form-factor PCs, industrial PCs |
| Application | Real-time edge inference |
Downloads
FAQ
How many cameras can 9 TOPS actually support?
The rated 1 channel is the baseline specification (1080P/25fps/1 algorithm per channel/2-second frame sampling). If you switch to 4-megapixel cameras, estimate around 2 channels; if you change to 1-second frame sampling, the number of supported channels is roughly halved; adding one more algorithm per channel halves it again.
What is the difference in compute usage between a 2-megapixel and a 4-megapixel camera?
The data volume of a single frame from a 4-megapixel camera is about twice that of a 2-megapixel camera, and the decoding and inference loads increase accordingly. With the same compute capacity, for 4-megapixel cameras it is recommended to plan for 50%–60% of the number of channels that 2-megapixel cameras can support, and to reserve headroom for peak loads.
How should I choose between 1-second and 2-second frame extraction?
For alarm-type scenarios (intrusion, fighting, smoke/fire, absence from post), 1 second is recommended—it gives faster response but doubles compute usage; for statistical scenarios (foot traffic, counting, on-duty duration), 2 seconds is sufficient and allows more channels. In mixed scenarios, you can set 1 second separately for key locations.
What algorithms does the device support?
They fall into 6 major categories by scenario: personnel behavior and on-duty status, safety risks and fire protection, vehicles and traffic, objects and facilities, environment and order, and food and kitchen, covering over 460 sub-scenarios. Which specific algorithms are enabled depends on the project configuration, and a list will be provided at delivery; algorithms involving human images must be used within the authorized scope.
Can we train the algorithm using our own on-site data?
Yes. The accompanying AI algorithm training platform supports four stages: data annotation, model training, model evaluation, and model release. You can import on-site images (either locally or via data回流), perform manual or intelligent annotation, and choose segmentation/detection/classification models for training. After evaluation passes, the model is quantized into an SG file and deployed to the device. The training platform requires a GPU server, which can be selected based on the project.
Both are 8 TOPS, so why can you support more channels?
Because the rated TOPS is only the peak upper limit; actual performance = peak × chip utilization. Common instruction set architectures (GPU/NPU) have a utilization of about 10%–30%, while reconfigurable dataflow architectures can reach 65%–95%. With the same 8 TOPS, the effective computing power of a dataflow solution is about 5.2–7.6 TOPS, so it can naturally support more channels; the specific number of channels is also affected by camera resolution, frame sampling interval, and the number of algorithms per channel.
Related Products
View All →
Shixinda AI’s AI Accelerator Card
48 TOPSOld Servers + Shixinda AI Core: Insert Shixinda AI accelerator card into existing servers to upgrade the computing power of monitoring centers without replacing…
Learn More →
Shixinda AI AI Accelerator Card (Edge Inference)
Legacy Devices + Shixinda AI Brain: Simply plug into servers, mini hosts, or industrial PCs to complete device-level AI upgrades.
Learn More →
SXD-X460H High-Performance Intelligent Computing Accelerator Card
New-Generation High-Performance AI Inference Accelerator Card, boasting a peak computing power of 48 TOPS, is widely compatible with mainstream servers featurin…
Learn More →Ready to go smart?
Leave a message or call us for tailored selection advice.






















