Products & compute
Modules, terminals, workstations, and servers cover loads from lightweight inference to high-density compute.
From one module
to a compute center
For algorithm inference, video compression, and intelligent IoT, GBOX provides an integrated hardware platform from front-end modules and edge terminals to workstations and compute servers.
One platform for full-scenario intelligent hardware
From product selection and chip adaptation to system delivery, unify hardware capability across endpoint, edge, and center.
Chips & architecture
Compatible with domestic chips and NVIDIA endpoint and data-center platforms, with a unified interface that lowers cross-platform adaptation and migration cost.
Industrial & security
Wide temperature, wide voltage, continuous operation, and domestic configurations for outdoor, industrial, and controllable deployments.
Delivery & services
Selection, algorithm porting, system integration, and on-site deployment shorten the path from validation to scaled rollout.
Core Hardware Product Matrix
Full coverage of edge terminals and edge hosts for one-stop multi-form hardware selection
Algorithm Inference Terminals
Embedded, light, standard, and high-compute tiers
Micro-compute inference terminal
AI compute0.5–5 TOPS
Ultra-compact, low-power design for front-end embedded scenarios with multi-channel lightweight inference and video structuring.
Micro-computeEmbeddedLow power
Light edge terminal
AI compute5–20 TOPS
Balanced compute and cost with concurrent multi-channel inference for small and medium edge analytics.
Mid-computeBoxMulti-channel
Standard edge terminal
AI compute20–80 TOPS
High-performance output for on-device large-model inference and multi-algorithm coordination.
High-computeLarge modelHigh density
High-compute edge terminal
AI compute80–300 TOPS
High compute density, active cooling, and quad-port design supporting NVIDIA terminal chips (Orin / Spark / Thor) and domestic high-compute chips.
Ultra-computeActive coolingQuad-portLarge model
Edge Compute Hosts
High-density compute · workstation and server dual form
Ultra-compute rack server (2U/3U)
AI compute300+ TOPS · configurable
2U/3U rack server supporting NVIDIA compute cards and domestic high-end chips for large-scale inference and high-concurrency cluster scheduling.
2U/3UCluster schedulingRack
High-performance workstation (air-cooled)
AI compute300+ TOPS · configurable
Portable/desktop workstation with air cooling for desktop deployment and mobile work, supporting domestic and general GPUs for local large-model inference and visualization.
Air coolingDesktopVisualization
High-performance workstation (liquid-cooled)
AI compute300+ TOPS · configurable
Liquid-cooled workstation with a transparent side showing the cooling system, supporting higher-power GPUs under sustained full load.
Liquid coolingHigh-power GPUSilent
Ultra-compute rack server (4U)
AI compute300+ TOPS · multi-card expansion
4U high-expansion rack server with multi-GPU full-height cards and large storage for massive model training, inference, and dense clusters.
4UMulti-GPUHigh expansion
Video Compression
H.265 / H.264 HD codec
Front-end embedded compression module
Multi-channel HD video compression with low bitrate and high quality in H.265/H.264, compact for front-end embedding.
H.265Low bitrateEmbedded
Video compression edge terminal
Edge-side multi-channel compression and transcoding for hundreds of channels, suited to monitoring centers and video aggregation.
Multi-channelEdge transcoding
High-density compression array server
Rack-mounted high-density compression array for large-channel compression, inference, and storage with integrated codec and transcoding.
RackLarge-channelInferenceStorage
Intelligent IoT Terminals
Vision · radar · depth · thermal multi-modal perception
Smart doorbell
Video intercom doorbell with face recognition, remote talk, and motion detection for home and office entrances.
Video intercomFaceMotion
Depth sensing module
ToF depth sensing for high-precision 3D perception, volume measurement, behavior detection, and spatial modeling.
ToF3D perception
Radar and thermal imaging
Millimeter-wave radar fused with infrared thermal imaging for all-weather monitoring, independent of light and weather.
mmWave radarThermal
IoT access gateway
Multi-protocol IoT gateway that unifies sensors and smart devices for edge aggregation and coordinated control.
Multi-protocolEdge aggregation
Face recognition panel
Wall-mounted vertical access panel with liveness, card, and password authentication for campuses and communities.
FaceLivenessAccess
Full-stack Chip Ecosystem
Deeply adapted to domestic and international AI chips for flexible compute selection
Multi-chip · unified architecture
The GBOX hardware ecosystem is deeply adapted to domestic and high-performance general-purpose chips with a unified hardware architecture and software interface. Choose by compute needs, budget, and localization requirements. The general-purpose route spans terminal to center, supporting NVIDIA terminal chips (Orin / Spark / Thor) and data-center compute cards.
Domestic camp
12 chip vendors · multi-tier compute adaptation
High-performance general-purpose chips
NVIDIA · terminal chips + compute cards
- Rockchip
- Kuxin
- Sophgo
- Cix
- Kunlunxin
- Denglin
- Houmo AI
- Huawei Ascend
- Hygon
- MetaX
- Iluvatar CoreX
- Moore Threads
- NVIDIA
Compute Matching Distribution
Terminal / host divide
TOPS
0.552080300Higher
Rockchip
Kuxin
Sophgo
Cix
Kunlunxin
Denglin
Houmo AI
Huawei Ascend
Hygon
MetaX
Iluvatar CoreX
Moore Threads
NVIDIA
Domestic chipsGeneral-purpose chipsTerminal / host divide (300 TOPS)
Supported chip vendors
Deeply adapted to mainstream AI chips for flexible compute selection
Domestic compute camp · 12 vendors
- Huawei Ascend
- Hygon
- Iluvatar CoreX
- MetaX
- Kunlunxin
- Houmo AI
- Moore Threads
- Denglin
- Sophgo
- Rockchip
- Cix
- Kuxin Microelectronics
High-performance general-purpose camp
- NVIDIA
More chips continuously adapting · stay tuned
Compute Selection Overview
From lightweight inference to high-density clusters, select by product form and workload.
Edge Terminal Series
Light / standard / high / ultra-compute tiers
- Low power, compact size, and dense compute output
- Low-power industrial design for wide temperature and voltage deployment
- For front-end perception, inference, video encoding, and analytics
- Embedded / box / wall-mount forms available
| Tier | Compute range | Representative chips | Typical scenarios |
|---|---|---|---|
| Micro-compute | 0.5 ~ 5 TOPS | Rockchip Kuxin | Front-end perception, lightweight inference, on-camera intelligence |
| Mid-compute | 5 ~ 20 TOPS | Sophgo Rockchip Cix | Access control, behavior analysis, structured extraction |
| High-compute | 20 ~ 80 TOPS | Kunlunxin Denglin Houmo AI NVIDIA terminal | Multi-channel video analytics, face recognition, complex algorithms |
| Ultra-compute | 80 ~ 300 TOPS | Huawei Ascend Hygon MetaX Iluvatar Moore Threads | Large-model inference, multi-algorithm coordination, regional aggregation |
Edge Host Series
For high-density inference, training, and cluster deployment
- Ultra-compute performance in rack server and workstation forms
- Rack / desktop / portable forms for server rooms and desks
- For large-scale aggregation analytics and high-concurrency inference scheduling
- Supports edge-cloud-edge coordination and unified multi-tier scheduling
| Tier | Form | Representative chips | Typical scenarios |
|---|---|---|---|
| Mid-compute | Entry server / workstation | Huawei Ascend NVIDIA terminal chips | Small/medium park aggregation, light workstation compute |
| High-compute | 2U/3U rack server / workstation | Hygon MetaX Iluvatar Moore Threads NVIDIA compute cards | Large-model inference, city-level aggregation, dense compute |
| Ultra-compute | 4U rack server / multi-card cluster | Huawei Ascend 910 Hygon NVIDIA high-end cards | Massive model inference, cluster scheduling |
Five-layer Software-hardware Architecture
From training platform to terminal perception, full-stack integration
01Layer 1 · Training
Algorithm training platform
Multi-framework compatible, distributed training, quantization optimized for all chip architectures
02Layer 2 · Operations
Algorithm operations platform
Full model lifecycle management, one-click rollout, inference monitoring, and compute scheduling
03Layer 3 · Platform
AIOT platform
Unified device access with lifecycle management for domestic and general protocols
04Layer 4 · Edge
Edge product layer
Edge terminals and edge hosts in dual form, covering lightweight inference, high-performance compute, and high-density cluster deployment.
05Layer 5 · Terminal
Terminal product layer
Vision, radar, depth, and thermal multi-modal perception for the front-end sensing network
Typical Scenario Hardware Combinations
End-to-end compute hardware combinations for smart security, smart transportation, smart manufacturing, public security, and more
Government and SOE parks
Domestic standard terminals + entry workstation, meeting Xinchuang and local data-processing needs.
TerminalStandard edge terminal, smart doorbellHostEntry workstation
Commercial industrial parks
Light front-end terminals + 2U aggregation server for unified security, staff management, and energy monitoring.
TerminalLight edge terminal, IoT gatewayHost2U rack server
City-level platforms
High-compute roadside terminals + 4U multi-card cluster in a three-tier architecture for city-scale data aggregation.
TerminalHigh-compute edge terminal, compression edge terminalHost4U rack server / multi-card cluster
Industrial sites
Micro-compute inspection terminals + high-performance workstation for defect detection, safety, and predictive maintenance.
TerminalMicro-compute inference terminal, embedded compression moduleHostHigh-performance workstation
Public security and judiciary
Fully domestic standard terminals + 2U/3U secure servers for smart policing, judicial supervision, and court hearings.
TerminalStandard edge terminal, depth sensing moduleHost2U/3U rack server
Core Technology and Custom Services
Full-process service from design to deployment
A professional team covers the full process from solution design to deployment. Whether hardware customization, algorithm porting, or system integration, we provide tailored solutions for your business needs.
Full-process technical supportFrom requirements assessment to deployment, 7×24 professional support
Hardware customization
Custom hardware development including form design, interface expansion, and compute configuration.
Algorithm porting
Rapid porting and optimization of mainstream frameworks, deeply tuned for each chip platform.
System integration
Software-hardware integration that connects edge-cloud-edge layers for unified data and management.
Technical support
7×24 professional response with remote and on-site support to keep systems running.
From one module to a compute center
Hand us your algorithms, channel counts, and site conditions for hardware and deployment advice.