- Server
- Components
- Carbon Fibre
- Analog to digital conversion chip
- power management IC
- storage
- H3C router
- H3C server
- Inspur server
- hard disk
- HART Loop Converter Analog Input
- reader
- Ethernet module
- EDGE Guardmaster
- MAT Guardmaster
- MatGuard Mat Manager
- relay
- power control
- switch
- servo drive
- I/O module
- Bridge connector (jumper plug)
- Human Machine Interface (HMI)
- power monitor
- 1420 power monitor
- 1411 current transformer
- socket
- system
- electric relay
- frequency converter
- Install the GND connector on the board
- transceiver
- Module Option Card
- module
0102030405
NVIDIA DGX Spark Super AI Computer
Product function and advantages
The function and advantages of NVIDIA DGX Spark product can be summarized in one sentence: it condenses the "data center level AI computing power" into a silent small box on the desktop, allowing individuals and small teams to develop, debug, and deploy large models like having a supercomputer. ）
1、 Product function (what can it help you with)
Local Large Model Development and Reasoning
A single machine can run 200 billion parameter level large language model prototypes, fine-tuning, and inference without relying on cloud GPUs.
Two units can infer ultra large scale models such as Llama-3.1-405B through NVLink-C2C interconnection.
A secure and compliant private AI environment
In scenarios where sensitive data such as hospitals, finance, and government cannot be uploaded to the cloud, full process AI research and development can be completed locally.
Teaching, Research, and Personal Experimentation
University courses, Kaggle competitions, paper reproduction, plug and play, saving server procurement and maintenance costs.
PoC (Proof of Concept) Accelerator
Small teams of enterprises quickly validate the feasibility of algorithms locally, and after passing the validation, smoothly migrate them to DGX Cloud or data centers for large-scale training, achieving zero code changes from desktop to cloud.
Edge AI and Robot Development
Run new generation AI models such as NVIDIA GR00T N1 robot base model and Cosmos Reason world model, directly driving robot prototypes or edge devices.
2、 Core advantages (why it's worth choosing)
Dimension specific advantages
Ultimate performance of 1000 TOPS (FP4)/1 PFLOPS AI computing power, reaching the level of previous generation data center GPU servers.
Extremely small volume of 150 × 150 × 50.5 mm, 1.2 kg, can be carried with one hand, standard socket power supply of 170 W, no need for cooling in the computer room.
Ultra large unified memory 128 GB LPDDR5X unified addressing, CPU/GPU sharing, avoiding frequent data copying and improving the efficiency of large models.
Ready to use pre installed Ubuntu 22.04 DGX OS, PyTorch/TensorFlow, CUDA-X, NVIDIA NIM, and other complete AI stacks, with optimized drivers, containers, and libraries.
Seamless expansion of ConnectX-7 200 GbE+NVLink-C2C, capable of horizontal scaling into multi machine clusters; One click migration of the same code to DGX Cloud without the need for refactoring.
Silent energy-saving passive cooling+silent fan, noise<35 dB, laboratory or office can operate for a long time.
Cost friendly pricing at the $3000 level, far lower than traditional GPU servers; Local operation saves a lot of cloud GPU rental costs.
Enterprise support can choose NVIDIA AI Enterprise subscription, which includes long-term technical support and certification drivers to meet commercial deployment SLA.
One sentence comparison:
The AI computing power that used to require a rack, several kilowatts of power consumption, and an operations team can now be placed on every developer's desktop with just a "dictionary sized box".
Application areas of the product
Although NVIDIA DGX Spark is a desktop level AI supercomputer, with 1 PFLOPS level computing power, 128 GB of unified memory, and 170 W power consumption, it can already independently run large models with 7-200 billion parameters locally. Therefore, it is particularly suitable for sinking large AI workflows that could only be placed in the cloud or data center to the "desktop/on-site/edge".
Below are typical application areas that can be directly implemented according to industry scenarios (★ represents the advantage of "single machine can run large models" that is particularly suitable for DGX Spark).
Electricity and Public Utilities
★ Real time scheduling of smart grid
Based on large-scale models for load forecasting and new energy output forecasting, a single machine can run a 70B parameter time-series forecasting network with 15 minute rolling prediction, replacing the traditional two-tier architecture of cloud training edge inference.
The on-site deployment is in the duty room of the dispatch center, and data cannot be sent out of the intranet to meet the safety and compliance of the power grid.
★ Equipment level predictive maintenance
Using a visual large-scale model to identify defects in visible light and infrared images of unmanned aerial vehicles/robots during inspections, it is possible to perform inference on 2000 4K images per second locally, achieving "flying and judging at the same time".
Perform multimodal fusion analysis of vibration partial discharge oil chromatography on high-voltage equipment such as transformers and GIS, and provide early warning of faults 3-6 months in advance.
★ Digital twin of power grid and emergency drill
Run a 1000 node level power grid flow simulation+reinforcement learning model to generate transfer strategies in minutes; Can quickly review extreme weather scenarios in the dispatch duty room or emergency command vehicle.
New energy and energy storage
★ Ultra short term prediction of wind and solar power
Single machine operation of 34 B meteorological power joint large model, integrating radar, satellite cloud images, and wind lidar data, updated every 15 minutes, reducing errors by 15-20%.
★ Energy storage EMS local strategy optimization
Run a multi-objective scheduling model (maximizing revenue, maximizing battery life, and controlling demand) locally in the energy storage station control room, and output charging and discharging strategies in seconds without the need to transmit them back to the cloud.
oil and gas
★ Earthquake interpretation and dessert prediction
Local 70 B parameter generative AI is used to automatically interpret faults and river sand bodies based on pre stack/post stack seismic data, generating three-dimensional sweet spot probability volumes. On site interpretation personnel can directly interact and modify them.
★ Prediction of drilling and completion risks
Using LSTM+Transformer hybrid model for real-time analysis of downhole while drilling data (MWD/LWD), warning of well kick and drilling jam risks 5-10 minutes in advance, and single machine can meet 20 Hz level real-time inference.
Nuclear power and thermal power
★ Security level AI monitoring
Run a 30 B parameter visual acoustic fusion model in the internal network of the nuclear power plant control room, analyze CCTV video and acoustic sensor data in real time, and discover leakage and abnormal equipment noise, meeting the physical isolation requirements of the nuclear power network.
★ Combustion optimization in thermal power plants
A combustion optimization strategy based on deep reinforcement learning, integrating real-time DCS data (temperature, pressure, oxygen content), single machine 100 ms closed-loop control, reducing coal consumption by 1-2%. Integrated energy/microgrid/industrial park
★ Campus level energy router brain
Deploy DGX Spark in the microgrid control room, run a 50B parameter multi energy flow optimization model, coordinate photovoltaic, energy storage, charging piles, and combined cooling, heating, and power supply in real time, and achieve "source grid load storage" second level autonomy.
★ Online carbon emissions accounting and trading
The local large-scale model reads system logs such as EMS, MES, ERP, etc., automatically generates minute level carbon emission reports, and directly connects to the regional carbon exchange API without data leaving the park. Summary in one sentence:
DGX Spark has truly brought the "big models that can only be placed in the cloud/computer room" to production lines such as power grid dispatch rooms, substation duty rooms, wind farm boosting stations, energy storage station control rooms, drilling platform offices, etc. for the first time, achieving "data not out of the field, models not in the cloud", greatly reducing the threshold and compliance risks of AI implementation in the energy and power industry.
Below are typical application areas that can be directly implemented according to industry scenarios (★ represents the advantage of "single machine can run large models" that is particularly suitable for DGX Spark).
Electricity and Public Utilities
★ Real time scheduling of smart grid
Based on large-scale models for load forecasting and new energy output forecasting, a single machine can run a 70B parameter time-series forecasting network with 15 minute rolling prediction, replacing the traditional two-tier architecture of cloud training edge inference.
The on-site deployment is in the duty room of the dispatch center, and data cannot be sent out of the intranet to meet the safety and compliance of the power grid.
★ Equipment level predictive maintenance
Using a visual large-scale model to identify defects in visible light and infrared images of unmanned aerial vehicles/robots during inspections, it is possible to perform inference on 2000 4K images per second locally, achieving "flying and judging at the same time".
Perform multimodal fusion analysis of vibration partial discharge oil chromatography on high-voltage equipment such as transformers and GIS, and provide early warning of faults 3-6 months in advance.
★ Digital twin of power grid and emergency drill
Run a 1000 node level power grid flow simulation+reinforcement learning model to generate transfer strategies in minutes; Can quickly review extreme weather scenarios in the dispatch duty room or emergency command vehicle.
New energy and energy storage
★ Ultra short term prediction of wind and solar power
Single machine operation of 34 B meteorological power joint large model, integrating radar, satellite cloud images, and wind lidar data, updated every 15 minutes, reducing errors by 15-20%.
★ Energy storage EMS local strategy optimization
Run a multi-objective scheduling model (maximizing revenue, maximizing battery life, and controlling demand) locally in the energy storage station control room, and output charging and discharging strategies in seconds without the need to transmit them back to the cloud.
oil and gas
★ Earthquake interpretation and dessert prediction
Local 70 B parameter generative AI is used to automatically interpret faults and river sand bodies based on pre stack/post stack seismic data, generating three-dimensional sweet spot probability volumes. On site interpretation personnel can directly interact and modify them.
★ Prediction of drilling and completion risks
Using LSTM+Transformer hybrid model for real-time analysis of downhole while drilling data (MWD/LWD), warning of well kick and drilling jam risks 5-10 minutes in advance, and single machine can meet 20 Hz level real-time inference.
Nuclear power and thermal power
★ Security level AI monitoring
Run a 30 B parameter visual acoustic fusion model in the internal network of the nuclear power plant control room, analyze CCTV video and acoustic sensor data in real time, and discover leakage and abnormal equipment noise, meeting the physical isolation requirements of the nuclear power network.
★ Combustion optimization in thermal power plants
A combustion optimization strategy based on deep reinforcement learning, integrating real-time DCS data (temperature, pressure, oxygen content), single machine 100 ms closed-loop control, reducing coal consumption by 1-2%. Integrated energy/microgrid/industrial park
★ Campus level energy router brain
Deploy DGX Spark in the microgrid control room, run a 50B parameter multi energy flow optimization model, coordinate photovoltaic, energy storage, charging piles, and combined cooling, heating, and power supply in real time, and achieve "source grid load storage" second level autonomy.
★ Online carbon emissions accounting and trading
The local large-scale model reads system logs such as EMS, MES, ERP, etc., automatically generates minute level carbon emission reports, and directly connects to the regional carbon exchange API without data leaving the park. Summary in one sentence:
DGX Spark has truly brought the "big models that can only be placed in the cloud/computer room" to production lines such as power grid dispatch rooms, substation duty rooms, wind farm boosting stations, energy storage station control rooms, drilling platform offices, etc. for the first time, achieving "data not out of the field, models not in the cloud", greatly reducing the threshold and compliance risks of AI implementation in the energy and power industry.
Product Operation Guide
NVIDIA DGX Spark Product Operation Guide
(Applicable to Retail/Founders Edition launched from July 2025 onwards) Unpacking and initial power on
Packing list: Host, 170W Gallium Nitride Power Adapter (including multi country plug), 1m USB-C to C cable, 1m 10 GbE Ethernet cable, Quick Start Card, Warranty Card.
Connection steps
① Insert the power adapter into the USB-C PD 3.1 port on the left side of the body;
② Connect HDMI 2.1 to a monitor or use USB-C to DP/HDMI cable to connect multiple screens;
③ Wired network: RJ-45 connected to 10 GbE; Wireless network: After turning on, select Wi Fi 7 according to the prompts;
④ Press the top power button for 1 second, and the indicator light changes from white to blue to indicate startup. The first startup takes about 90 seconds to enter the DGX OS desktop.
System and account initialization
Default system: Ubuntu 22.04 LTS+NVIDIA DGX OS 5.2 (kernel 6.8+CUDA 12.6). First Launch Wizard
① Language → ② Time Zone → ③ Create a local administrator account (with sudo privileges) → ④ Choose whether to register for NVIDIA AI Enterprise (90 day trial).
Remote SSH: The system has pre installed openssh server; Execute after first networking
sudo systemctl enable --now ssh
default port
(Applicable to Retail/Founders Edition launched from July 2025 onwards) Unpacking and initial power on
Packing list: Host, 170W Gallium Nitride Power Adapter (including multi country plug), 1m USB-C to C cable, 1m 10 GbE Ethernet cable, Quick Start Card, Warranty Card.
Connection steps
① Insert the power adapter into the USB-C PD 3.1 port on the left side of the body;
② Connect HDMI 2.1 to a monitor or use USB-C to DP/HDMI cable to connect multiple screens;
③ Wired network: RJ-45 connected to 10 GbE; Wireless network: After turning on, select Wi Fi 7 according to the prompts;
④ Press the top power button for 1 second, and the indicator light changes from white to blue to indicate startup. The first startup takes about 90 seconds to enter the DGX OS desktop.
System and account initialization
Default system: Ubuntu 22.04 LTS+NVIDIA DGX OS 5.2 (kernel 6.8+CUDA 12.6). First Launch Wizard
① Language → ② Time Zone → ③ Create a local administrator account (with sudo privileges) → ④ Choose whether to register for NVIDIA AI Enterprise (90 day trial).
Remote SSH: The system has pre installed openssh server; Execute after first networking
sudo systemctl enable --now ssh
default port
Product after-sales and maintenance
NVIDIA DGX Spark Super AI Computer
Overview of one-stop after-sales and maintenance services
(Applicable to the retail version officially shipped in Chinese Mainland from July 2025)
Official warranty policy
|Warranty period | 3-year limited warranty for the entire machine (including the host, power adapter, and accompanying cables) IT Home|
|Warranty Start | Based on the invoice date or factory date plus 30 days, whichever is earlier|
|Warranty coverage | Hardware failures, original firmware defects, pre installed software crashes|
|Non warranty scope | Human damage, water ingress, lightning strikes, self disassembly, unauthorized firmware flashing
Overview of one-stop after-sales and maintenance services
(Applicable to the retail version officially shipped in Chinese Mainland from July 2025)
Official warranty policy
|Warranty period | 3-year limited warranty for the entire machine (including the host, power adapter, and accompanying cables) IT Home|
|Warranty Start | Based on the invoice date or factory date plus 30 days, whichever is earlier|
|Warranty coverage | Hardware failures, original firmware defects, pre installed software crashes|
|Non warranty scope | Human damage, water ingress, lightning strikes, self disassembly, unauthorized firmware flashing
description2