NVIDIA DGX Spark Personal AI Supercomputer GB10 Grace Blackwell - 128GB / 4TB
PRODUCT DESCRIPTION
NVIDIA DGX Spark Personal AI Supercomputer, NVIDIA GB10 Grace Blackwell Superchip, 20-Core Arm CPU, 128GB Unified Memory, 4TB Self-Encrypting NVMe SSD, ConnectX-7 200GbE, Wi-Fi 7 (940-54242-0007-000) - 1 Year Local Warranty
NVIDIA DGX Spark β NVIDIA's Own Desktop AI Supercomputer, Built for Local LLMs and Autonomous Agents
The NVIDIA DGX Spark (part number 940-54242-0007-000) is NVIDIA's first-party build of the DGX Spark platform: a 150 Γ 150 mm, 1.2 kg desktop AI supercomputer powered by the NVIDIA GB10 Grace Blackwell Superchip. It pairs a 20-core Arm CPU with a Blackwell-architecture GPU and 128 GB of coherent unified LPDDR5x memory, delivering up to 1 petaFLOP (1,000 TOPS) of FP4 AI performance β enough to run inference on large language models of up to 200 billion parameters and fine-tune models of up to 70 billion parameters entirely on your desk, with no cloud GPU rental, no data leaving your premises and no per-token cost.
NVIDIA positions DGX Spark as a desktop agent computer: large local memory, always-on efficiency and the full NVIDIA AI software stack preinstalled on DGX OS make it the reference platform for building and running autonomous AI agents locally. This part number ships factory-sealed with a US power cord; SourceIT supplies a UK/Singapore Type-G power cord separately with every unit, so it is ready to plug in on arrival. SourceIT, an authorised IT reseller in Singapore, supplies it with a full GST invoice, 1 year local warranty and enterprise procurement support.
NVIDIA GB10 Grace Blackwell Superchip β 1 PetaFLOP of FP4 AI Compute
The GB10 Superchip combines a 20-core Arm CPU (10 Cortex-X925 performance cores and 10 Cortex-A725 efficiency cores) with an NVIDIA Blackwell GPU on a single package, connected by NVLink-C2C so CPU and GPU share one memory space rather than copying data across a PCIe bus. Fifth-generation Tensor Cores with native FP4 and FP8 precision drive up to 1 PFLOP of sparse FP4 throughput for LLM inference and fine-tuning; fourth-generation RT Cores accelerate rendering, simulation and synthetic-data pipelines; and dedicated NVENC and NVDEC engines handle video encode and decode for computer-vision and multimodal workloads. The whole superchip runs at a 140 W TDP.
128 GB Coherent Unified Memory at 273 GB/s β Models up to 200 Billion Parameters
Conventional workstations hit a VRAM wall long before a 70B or 120B model loads. DGX Spark's 128 GB of LPDDR5x is a single coherent pool shared by CPU and GPU over a 256-bit, 273 GB/s interface, so models such as Llama, Qwen, DeepSeek, Gemma, Mistral, gpt-oss and FLUX-class diffusion models load in full. In practical terms: a single unit runs inference on models up to 200 billion parameters (FP4), fine-tunes models up to 70 billion parameters, and gives data scientists a 128 GB workspace for RAPIDS, embeddings, vector search and large-dataset analytics without paging to disk.
4 TB Self-Encrypting NVMe SSD, ConnectX-7 200 GbE, Wi-Fi 7 and Four USB-C Ports
This configuration ships with a 4 TB NVMe M.2 self-encrypting SSD β room for multiple multi-hundred-gigabyte model checkpoints, datasets and container images. Networking is where DGX Spark separates itself from every other mini PC: an integrated NVIDIA ConnectX-7 SmartNIC provides 200 Gbps RDMA-capable connectivity for linking units together or attaching to fast storage, alongside a 10 GbE RJ-45 port for the office network and Wi-Fi 7 with Bluetooth 5.4 for cable-free desk placement. Four USB Type-C ports, HDMI 2.1a with multichannel audio and up to three DisplayPort outputs over USB-C Alt Mode cover peripherals and monitors.
Scale to 405B and 700B Parameters by Linking DGX Spark Units
Two DGX Spark units connected over ConnectX-7 with a QSFP112 direct-attach cable work as one system for models of up to 405 billion parameters, and NVIDIA now supports clustering up to four DGX Spark systems to work with models of up to 700 billion parameters. That lets a Singapore team start with one desktop unit and grow into a small private AI cluster without a server room β pair this listing with the NVIDIA QSFP112 400G DAC cable (X0101G00400A), in stock at SourceIT.
NVIDIA DGX OS, NVIDIA AI Enterprise, NemoClaw and the Full NVIDIA AI Software Stack Preinstalled
DGX Spark boots into NVIDIA DGX OS, an Ubuntu-based Linux build with the NVIDIA AI stack ready on day one: CUDA, cuDNN, TensorRT and TensorRT-LLM, NVIDIA NIM microservices, NVIDIA NeMo, Triton Inference Server, RAPIDS, NGC-certified containers, PyTorch, TensorFlow, JAX, vLLM, Ollama and Hugging Face tooling. NVIDIA AI Enterprise for DGX Spark adds enterprise-grade security, optimised performance and support for production agentic AI. The 2026 DGX OS updates added streamlined NemoClaw installation β NVIDIA's open-source reference stack that wraps OpenClaw agents with OpenShell security and privacy guardrails β plus the latest open models and up to 1.9Γ faster inference, on top of the roughly 2Γ performance uplift NVIDIA delivered in its January 2026 software release. Because this is the same software foundation as DGX Station and DGX data-centre systems, work prototyped here moves to larger NVIDIA infrastructure or DGX Cloud without code changes.
Quiet, 240 W, 1.2 kg β Runs on a Desk, Not in a Data Centre
A 240 W external power adapter from a standard wall socket is all DGX Spark needs β no dedicated circuit, no rack, no liquid cooling. NVIDIA's declared ECMA-109 acoustic figures are a 35 dB sound-power level under maximum GPU stress and 19 dB at idle, so it sits comfortably next to a monitor in an open office, research lab or home office. UEFI Secure Boot, TPM-based security and the self-encrypting SSD keep models and data protected at rest.
Available in Singapore Through SourceIT
SourceIT is an authorised IT reseller in Singapore and supplies the NVIDIA DGX Spark with a full GST invoice, 1 year local warranty and enterprise, research-grant and education procurement assistance. This 940-54242-0007-000 unit ships factory-sealed with a US power cord and SourceIT includes a separate UK/Singapore Type-G cord; if you prefer the factory-bundled UK-cord variant, see the NVIDIA DGX Spark Founders Edition (940-54242-0009-000). Read our NVIDIA DGX Spark Setup and Enterprise Deployment Guide before your unit arrives, and browse every GB10 and GB300 system we stock in the NVIDIA DGX Spark & DGX Station collection.
NVIDIA DGX Spark Datasheet β Technical Specifications
General
| Product Name | NVIDIA DGX Spark AI Supercomputer β 4TB |
| Part Number | 940-54242-0007-000 (GTIN 812674029197) |
| Brand / Build | NVIDIA first-party (NVIDIA-built DGX Spark, Founders Edition hardware) |
| Category | Personal AI supercomputer / desktop agent computer / AI developer workstation |
| Form Factor | Ultra-compact desktop, 150 Γ 150 mm footprint |
| Local Warranty (Singapore) | 1 Year local warranty; procurement and GST invoice by SourceIT |
Processor & GPU β NVIDIA GB10 Grace Blackwell Superchip
| Architecture | NVIDIA Grace Blackwell β CPU + GPU on one superchip, NVLink-C2C interconnect |
| CPU | 20-core Arm: 10 Γ Cortex-X925 performance cores + 10 Γ Cortex-A725 efficiency cores |
| GPU | NVIDIA Blackwell architecture, Blackwell-generation CUDA cores |
| Tensor Cores | 5th generation, FP4 / FP8 optimised |
| RT Cores | 4th generation |
| Video Engines | 1 Γ NVENC, 1 Γ NVDEC |
| GB10 TDP | 140 W (CPU + GPU) |
AI Performance
| Peak AI Compute | Up to 1 PFLOP (1,000 TOPS) FP4 Tensor performance with sparsity |
| Inference | AI models up to 200 billion parameters on a single unit |
| Fine-Tuning | AI models up to 70 billion parameters on a single unit |
| 2 Linked Units | AI models up to 405 billion parameters over ConnectX-7 |
| Up to 4 Linked Units | AI models up to 700 billion parameters, per NVIDIA |
Memory
| System Memory | 128 GB LPDDR5x coherent unified system memory, shared by CPU and GPU |
| Memory Interface | 256-bit |
| Memory Bandwidth | 273 GB/s |
| Upgradeability | Soldered β not user-upgradeable |
Storage
| SSD | 4 TB NVMe M.2 with self-encryption (SED) |
Networking
| High-Speed NIC | NVIDIA ConnectX-7 SmartNIC @ 200 Gbps, RDMA capable, for multi-node clustering |
| Ethernet | 1 Γ RJ-45, 10 GbE |
| Wi-Fi | Wi-Fi 7 (802.11be) |
| Bluetooth | Bluetooth 5.4 |
I/O & Display
| USB | 4 Γ USB Type-C |
| Display Outputs | 1 Γ HDMI 2.1a; up to 3 Γ DisplayPort over USB-C (DP Alt Mode) |
| Audio | HDMI multichannel audio output |
Software & Security
| Operating System | NVIDIA DGX OS (Ubuntu Linux based), arm64 |
| AI Stack (preinstalled) | CUDA, cuDNN, TensorRT / TensorRT-LLM, NVIDIA NIM, NVIDIA NeMo, Triton, RAPIDS, NGC containers, PyTorch, TensorFlow, JAX, vLLM, Ollama |
| Agentic AI | NVIDIA Agent Toolkit with OpenShell; NemoClaw reference stack via DGX OS update; DGX Spark Playbooks |
| Enterprise Software | NVIDIA AI Enterprise β DGX Spark (evaluation licence available from NVIDIA) |
| Security | UEFI Secure Boot, TPM-based security, self-encrypting SSD |
Power, Physical & Acoustics
| Power Supply | 240 W external power adapter |
| Power Cord | US Type-B cord in the factory-sealed box; UK/Singapore Type-G cord supplied separately by SourceIT |
| Dimensions (L Γ W Γ H) | 150 mm Γ 150 mm Γ 50.5 mm |
| Weight | 1.2 kg |
| Declared Noise (ECMA-109) | Sound power LWA,m 35 dB operating (max GPU stress, 25 Β°C) / 19 dB idle; emission sound pressure 29 dB / 13 dB |
Compatibility & Software Support
What Runs on DGX Spark
Anything in the mainstream NVIDIA and open-source AI ecosystem that is published for arm64: Hugging Face Transformers and Diffusers, vLLM, Ollama, llama.cpp, LM Studio-style local inference servers, LangChain and LlamaIndex RAG pipelines, NVIDIA NIM microservices, NeMo for fine-tuning and guardrails, TensorRT-LLM for optimised serving, ComfyUI and FLUX-class image generation, RAPIDS cuDF/cuML for data science, and the NVIDIA Isaac, Metropolis and Holoscan frameworks for robotics, vision and edge applications. NVIDIA publishes step-by-step DGX Spark Playbooks covering local LLM serving, fine-tuning, RAG, agents and vision workloads so new users are productive within the first hour.
Arm64 Architecture β Plan Your Software Accordingly
The GB10's CPU is Arm-based, so DGX OS, drivers and container images are arm64 (aarch64) builds. NVIDIA NGC containers and every major AI framework ship arm64 builds, but proprietary x86-only desktop applications will not run natively, and Windows is not supported. Teams that need a Windows desktop alongside Spark typically keep a laptop and use the Spark over SSH, Jupyter, VS Code Remote or the DGX Dashboard.
Clustering and Networking
The onboard ConnectX-7 200 Gbps SmartNIC links two units with a single QSFP112 direct-attach cable for 405B-parameter models, and NVIDIA supports up to four units for 700B-parameter models. The same NIC supports RDMA to NVMe-over-fabric storage. The 10 GbE RJ-45 port handles site networking, and Wi-Fi 7 keeps the unit connected on a desk without cabling.
NVIDIA GB10 Platform Siblings
DGX Spark is the reference design behind the GB10 desktops from ASUS (Ascent GX10), Dell (Pro Max with GB10), Lenovo (ThinkStation PGX), HP (ZGX Nano), GIGABYTE (AI TOP ATOM), MSI (EdgeXpert), Acer (Altos BrainSphere GB10) and xFusion (FusionXpark GB10). All share the GB10 Superchip, 128 GB unified memory and 1 PFLOP FP4; they differ in chassis, storage tier, warranty term and price. The NVIDIA-built unit on this page is the first-party gold-finish Founders Edition hardware with 4 TB storage and the ConnectX-7 SmartNIC fitted as standard.
Package Contents & Ordering Information
What Ships in the Box
- NVIDIA DGX Spark unit with NVIDIA GB10 Grace Blackwell Superchip, 128 GB unified memory and 4 TB self-encrypting NVMe SSD
- 240 W power adapter
- US power cord (factory-sealed), plus a UK/Singapore Type-G power cord supplied separately by SourceIT
- NVIDIA DGX OS preinstalled with the NVIDIA AI software stack
- NVIDIA quick start documentation
The QSFP112 direct-attach cable for linking two units is sold separately β see the NVIDIA QSFP112 400G DAC cable 40 cm (X0101G00400A).
NVIDIA DGX Spark & GB10 Systems at SourceIT Singapore
The NVIDIA DGX Spark 4TB (940-54242-0007-000) is supplied by SourceIT Singapore with GST invoice and 1 year local warranty.
- NVIDIA DGX Spark Founders Edition 4TB (940-54242-0009-000, UK cord bundled) β the same NVIDIA-built DGX Spark 4TB; differs only in the bundled power cord
- ASUS Ascent GX10 β GB10 platform in 1TB, 2TB and 4TB tiers
- Dell Pro Max with GB10 β GB10 platform with 3-year onsite warranty, 2TB and 4TB
- Lenovo ThinkStation PGX β GB10 platform, 1TB and 4TB
- HP ZGX Nano G1n β GB10 platform, 4TB
- GIGABYTE AI TOP ATOM β GB10 platform, 1TB and 4TB
- MSI EdgeXpert MS-C931 β GB10 platform, 4TB
- Acer Altos BrainSphere GB10 F1 β GB10 platform, 4TB
- xFusion FusionXpark GB10 β GB10 platform with bundled technical support, 1TB / 2TB / 4TB
- MSI XpertStation WS300 (NVIDIA DGX Station GB300) β step up to 748 GB coherent memory and 20 PFLOPS for trillion-parameter models
Browse the full NVIDIA DGX Spark & DGX Station collection. For volume, project, research-grant or education pricing, GST invoicing and current lead times, email sales@sourceit.com.sg.
Datasheet, Downloads & Setup Guide
Official NVIDIA documentation and resources for DGX Spark:
- NVIDIA DGX Spark β official product page and specifications (Singapore)
- NVIDIA DGX Spark datasheet (PDF)
- NVIDIA DGX Spark Quick Start Guide (PDF)
- NVIDIA DGX Spark User Guide and documentation
- NVIDIA DGX Spark Playbooks β step-by-step recipes for LLMs, fine-tuning, RAG and agents
- NVIDIA DGX Spark / GB10 developer forum and FAQ
- NVIDIA DGX Spark Setup and Enterprise Deployment Guide β SourceIT
Frequently Asked Questions β NVIDIA DGX Spark
What is the NVIDIA DGX Spark?
A personal AI supercomputer the size of a small book: NVIDIA's GB10 Grace Blackwell Superchip, 128 GB of coherent unified memory and up to 1 PFLOP of FP4 AI performance in a 150 Γ 150 mm, 1.2 kg desktop unit. NVIDIA describes it as a desktop agent computer β a complete platform for developing, running and serving large language models and autonomous AI agents locally.
What is the difference between part numbers 940-54242-0007-000 and 940-54242-0009-000?
Identical NVIDIA-built DGX Spark hardware β GB10 Grace Blackwell, 128 GB unified memory, 4 TB self-encrypting SSD, ConnectX-7. The 0007-000 on this page ships with a US power cord (SourceIT adds a UK/Singapore cord separately); the 0009-000 Founders Edition ships factory-bundled with a UK Type-G cord.
What AI models can a single DGX Spark run?
Inference on models up to 200 billion parameters and fine-tuning of models up to 70 billion parameters β in practice Llama, Qwen, DeepSeek, Gemma, Mistral and gpt-oss-class LLMs, vision-language models and FLUX-class image generation, running from the 128 GB unified memory. Two linked units reach 405 billion parameters; up to four reach 700 billion.
How does DGX Spark compare to the ASUS, Dell, Lenovo, HP, GIGABYTE, MSI, Acer and xFusion GB10 desktops?
They all share the NVIDIA GB10 Superchip, 128 GB unified memory and 1 PFLOP FP4, so raw AI performance is the same. DGX Spark is NVIDIA's own first-party Founders Edition hardware with 4 TB self-encrypting storage and the ConnectX-7 SmartNIC as standard; the OEM units differ in chassis design, storage tiers (1TB / 2TB / 4TB), warranty term (for example Dell's 3-year onsite) and price. SourceIT stocks every one of them, so you can choose on budget and support model.
How does DGX Spark compare to the NVIDIA DGX Station GB300?
Different class. DGX Spark is a desktop-scale GB10 system for individual developers, researchers and small teams working with models up to about 200 billion parameters. DGX Station GB300 systems (MSI XpertStation WS300, ASUS ET900N G3, HP ZGX Fury G1n, GIGABYTE W775) carry 748 GB of coherent memory and up to 20 PFLOPS for trillion-parameter models and team-scale production AI.
Can I link two or more DGX Spark units together?
Yes. The built-in ConnectX-7 200 Gbps SmartNIC connects two units with a QSFP112 direct-attach cable (sold separately) to work as one system on models up to 405 billion parameters, and NVIDIA supports clustering up to four units for models up to 700 billion parameters.
What operating system and software come preinstalled?
NVIDIA DGX OS, an Ubuntu-based Linux build with the NVIDIA AI software stack ready to use: CUDA, TensorRT, NVIDIA NIM, NeMo, Triton, RAPIDS, NGC containers, PyTorch, TensorFlow, vLLM and Ollama. Recent DGX OS updates add NemoClaw for secure always-on agents and significant inference speed-ups. NVIDIA AI Enterprise for DGX Spark is available for production deployments.
Does DGX Spark run Windows or x86 software?
No. The Grace CPU is Arm64, so DGX Spark runs arm64 Linux software and containers only. Every mainstream AI framework and NVIDIA NGC container is published for arm64, but proprietary x86-only desktop applications will not run natively.
Does it need special power, cooling or a server room?
No. DGX Spark runs from a 240 W external adapter on a standard wall socket, is actively air-cooled, and NVIDIA's declared noise figures are 35 dB under full GPU load and 19 dB at idle β quiet enough for an open-plan office or lab bench. This part number's factory box contains a US cord; SourceIT supplies the UK/Singapore cord with it.
Who is DGX Spark for?
AI developers and ML engineers prototyping and fine-tuning models before scaling to DGX Cloud or data-centre GPUs; researchers and university labs that need on-premise compute; data scientists running large-memory analytics; robotics, smart-city and computer-vision teams building on Isaac, Metropolis and Holoscan; and enterprises in finance, healthcare, legal, defence and government that must keep data and models on-premise for privacy and compliance.
Is the DGX Spark covered by local warranty in Singapore?
Yes. Every NVIDIA DGX Spark sold by SourceIT comes with 1 year local warranty and a full GST invoice. Volume, enterprise, research-grant and education procurement enquiries are welcome.
How do I buy the NVIDIA DGX Spark in Singapore?
Order online on this page or email sales@sourceit.com.sg quoting part number 940-54242-0007-000. SourceIT is an authorised IT reseller in Singapore and supplies DGX Spark with GST invoice, local warranty, the QSFP112 cable for clustering and the rest of the NVIDIA DGX Spark & DGX Station range.
