Skip to main content
SHIXINDA 时信达
SXD-X3 9 TOPS edge inference accelerator card, product image
Accelerator Card

Shixinda AI AI Accelerator Card (Edge Inference)

Specification · PCIe Card · Servers, Mini Hosts, Industrial Control PCs

Legacy Devices + Shixinda AI Brain: Simply plug into servers, mini hosts, or industrial PCs to complete device-level AI upgrades.

Item
Specification
Interface Form
PCIe Card
Compatible Devices
Servers, Mini Hosts, Industrial Control PCs
Application Direction
Edge-side Real-time AI Inference

No. shixinda-ai-ai-accelerator-card-edge-inference · 2026-11-19

Overview

Product Overview: Legacy Devices + AI Brain

This Shixinda AI Accelerator Card (for edge inference) is designed for computing power upgrading of existing devices. Deployed devices including servers, small form factor hosts, industrial control units (ICUs) do not require full replacement; simply inserting this accelerator card enables edge AI inference capability, supporting real-time intelligent identification applications. It is the most cost-effective and fastest-to-deploy device-level AI transformation solution.

Upgrade & Transformation Value

  • Leverage Existing Devices: Upgrade existing servers, small form factor hosts, industrial control units (ICUs) directly via card insertion
  • Cost Controllability: High computing power cost-effectiveness, enabling real-time inference capability with minimal investment
  • Edge Deployment: Local inference close to data sources, with low latency and bandwidth saving

Core Features

  • High Computing Power Cost-Effectiveness: Delivers sufficient edge inference computing power at low cost
  • Multi-Core AI Inference Engine: Supports real-time AI inference
  • Wide Compatibility: Compatible with device forms including servers, small form factor hosts, industrial control units (ICUs), etc.

Typical Transformation Scenarios

Edge AI upgrading scenarios for legacy devices such as access control turnstiles, security inspection equipment, industrial inspection systems, etc.

Capacity and channel definition

The quoted channel count refers to live video channels under a baseline of 1080P, 25 fps, one algorithm per channel and one sampled frame every 2 seconds. Real capacity varies with three factors:

VariableChangeEffect on compute
Camera resolution2 MP → 4 MPDecode and inference load per channel rises by 50%–100%; channel count drops
Sampling interval2 s → 1 sFrames processed per second double; channel count roughly halves
Algorithms per channel1 → 2Inference calls double; channel count roughly halves

Rule of thumb: convert by "resolution × sampling rate × algorithms per channel", then keep 30% headroom. 2 MP with 2-second sampling and a single algorithm corresponds to the quoted channel count; for 4 MP or 1-second sampling, plan for 50%–60% of it.

Dataflow architecture and effective compute

The nominal TOPS is only an upper bound — what determines real capacity is chip utilization. This device runs on a reconfigurable dataflow AI chip: the whole model is compiled into a data pipeline, so operands flow through the hardware and computation completes on the way, without repeatedly fetching instructions or moving intermediate results in and out of memory.

Chart comparing chip utilization: instruction-set architectures 10–30% versus reconfigurable dataflow 65–95%; 8 TOPS translates to 0.8–2.4 TOPS versus 5.2–7.6 TOPS

ItemInstruction-set (typical GPU / NPU)Dataflow (this device)
Executionoperators run one by one, intermediates round-trip through memoryoperator graph compiled into a pipeline
Chip utilization (public comparison)10%–30%65%–95%
Effective compute at 8 TOPSabout 0.8–2.4 TOPSabout 5.2–7.6 TOPS
Channels at equal nominal TOPSbaselineabout 2–3× by median utilization
Model deploymentframework adaptation and manual tuningin-house compiler, TensorFlow / PyTorch / ONNX compiled automatically

Effective compute = peak TOPS × utilization. Three practical consequences: fewer devices for legacy upgrades (saving rack space and power), algorithms can be recompiled and redeployed without replacing hardware, and utilization stays high under concurrent decoding plus inference.

Utilization figures are publicly reported architecture ranges; measured values vary with operator structure, input resolution and batching strategy.

Algorithms

Algorithms are grouped into six scenario families, covering 460+ fine-grained scenarios. Which ones are enabled is selected per project and confirmed in the delivery list.

Scenario familyRepresentative algorithmsTypical locations
People and attendanceintrusion, line crossing, loitering, lingering, climbing, fall, on-duty status, PPE compliance, helmet / vest / uniform / mask / glove detectionperimeter, guard room, work area
Safety and firesmoke, open flame, fire lane obstruction, restricted-area intrusion, blocked passage, gas leak, static dischargewarehouse, switch room, kitchen, hazardous area
Vehicles and trafficplate recognition, illegal parking, wrong-way driving, traffic counting, non-motor vehicle intrusion, engineering vehicles, height limitcampus gate, site entrance, road
Objects and facilitiesmaterial stacking, pallets, shelf gaps, equipment status, gauge reading, conveyor anomalyworkshop, warehouse, power room
Environment and orderbin overflow, standing water, dust, street occupation, cluttersite roads, waste room, construction area
Food and kitchenkitchen uniform, rodent detection, sample retention, tableware sanitising, cold-chain temperatureschool and staff canteens

Alert data flow

Shixinda AI alert data flow: camera streaming, edge decode and inference, local alarm, report to platform, dedup and push, operator action — about 5 to 8 seconds end to end

Deployment architecture

Shixinda AI video analytics deployment architecture: capture, edge analytics, central platform and applications, with standalone, cluster and large-cluster models

Algorithm training & customization

Shixinda AI also provides an algorithm training platform: annotate site data, train segmentation / detection / classification models, evaluate, then publish a quantized SG file and load it onto the edge devices.

Shixinda AI algorithm training workflow: data collection, annotation, training, evaluation, publishing and deployment, with training environment and continuous improvement loop

Related

References

Related Chinese regulations and standards for further reference:

Applications

工厂
工厂
化工
化工
矿山
矿山
电力
电力

Core Features

To be added

This section is pending content and will appear once filled in the admin.

To be added

This section is pending content and will appear once filled in the admin.

To be added

This section is pending content and will appear once filled in the admin.

To be added

This section is pending content and will appear once filled in the admin.

Specifications

Hardware Parameters

ItemSpecification
Interface FormPCIe Card
Compatible DevicesServers, Mini Hosts, Industrial Control PCs
Application DirectionEdge-side Real-time AI Inference
标签:General

FAQ

Does SXD-X3 work with our existing cameras?

Yes. It connects to existing IP cameras over RTSP or GB/T 28181 and keeps your current cameras and NVR, so no camera replacement is needed.

How many channels can SXD-X3 handle?

The rated capacity is %s. Actual capacity changes with camera resolution, frame sampling interval and the number of algorithms per channel.

What is included in delivery?

Supply, on-site installation and commissioning, platform integration and operator training, with remote support afterwards.

Ready to go smart?

Leave a message or call us for tailored selection advice.

Call Hotline
Call