CtrlK
Template:Infobox product
NVIDIA DGX Spark is a compact, deskside AI development system in the DGX product family created by NVIDIA. First announced as "Project DIGITS" at CES 2025 and officially launched on 15 October 2025, DGX Spark places a Grace Blackwell Superchip-based AI computer on a desktop, targeting developers, researchers, and students who need local fine-tuning and inference without relying exclusively on the cloud.[1][2]
DGX Spark is powered by the GB10 Grace Blackwell Superchip and is specified by NVIDIA for up to one FP4 petaFLOP of AI performance (theoretical, using sparsity). It features 128 GB of coherent unified memory, a 4 TB self-encrypting NVMe M.2 SSD, and a ConnectX-7 NIC that enables low-latency peer-to-peer links between two units. NVIDIA positions Spark for prototyping, local fine-tuning (up to ~70 billion parameters), and inference with models up to ~200 billion parameters.[3][4]
DGX Spark originated as "Project DIGITS," a small form-factor AI computer first announced at CES 2025 in January and previewed at GTC 2025 in March. Initially targeted for a May 2025 release, the launch was delayed to October 2025 due to production and supply chain adjustments.[5] Coverage at the time highlighted a preorder price change from US$3,000 for the concept to US$3,999 for the production unit and the introduction of a larger sibling, DGX Station, based on the GB300 Superchip.[2][6]
NVIDIA formally announced that DGX Spark systems would begin shipping the week of 13 October 2025, with sales starting on October 15, 2025. Availability was through NVIDIA.com and retail partners including Micro Center, with OEMs introducing GB10-based alternatives.[1][7]
DGX Spark uses the NVIDIA GB10 Grace Blackwell Superchip, a system-on-a-chip (SoC) manufactured using TSMC's 3nm process technology with 208 billion transistors.[8] The chip consists of two dielets in a 2.5D package:
S-Dielet: Contains the CPU and memory subsystem
G-Dielet: Houses the GPU
The superchip integrates a Blackwell-generation GPU and a 20-core Arm CPU on one package, connected via NVLink-C2C at 600 GB/s and sharing a unified LPDDR5x memory pool. MediaTek stated it co-designed aspects of the GB10 Superchip used in DGX Spark.[9]
The platform specifications include:
Up to 1 PFLOP (theoretical) at FP4 precision using 5th-generation Tensor Cores
31 TFLOPS FP32, 1000 TOPS FP4[8]
128 GB coherent unified LPDDR5x memory (256-bit interface, ~273 GB/s bandwidth)[10]
4 TB self-encrypting NVMe M.2 storage in the base configuration[4]
I/O: 4× USB-C (one used for power), 1× HDMI 2.1a, 1× RJ-45 10 GbE, Wi-Fi 7, Bluetooth 5.4[10]
TDP: 140W[8]
Independent hands-on reviews have reported the CPU core makeup (10 Cortex-X925 "performance" + 10 Cortex-A725 "efficiency" cores) and confirmed the compact metal enclosure and USB-C power input (external PSU).[11][12]
The system features 128 GB of LPDDR5x coherent unified memory. This architecture allows both the CPU and GPU to access the same memory pool seamlessly, eliminating the need for data transfers between system RAM and GPU VRAM. This is particularly beneficial for large AI models that exceed the capacity of typical discrete GPU memory.[11]
Spark includes an onboard ConnectX-7 Smart NIC with QSFP ports. NVIDIA's specifications list the NIC at up to 200 Gb/s and enable linking two DGX Spark systems directly; NVIDIA markets this as supporting model execution up to ~405 billion parameters across a two-node setup.[1][3][13]
DGX Spark ships with the NVIDIA AI software stack and DGX OS preinstalled, an Ubuntu-based operating system (Ubuntu 24.04 with Linux 6.11 kernel) optimized for AI workloads.[14] The software ecosystem includes:
CUDA libraries and tooling for local development
NVIDIA NIM (NVIDIA Inference Microservices)
Docker and NVIDIA Container Toolkit support
Access to NVIDIA NGC catalog for pre-trained models
Support for standard frameworks like PyTorch, TensorFlow, and JAX
NVIDIA NeMo for model fine-tuning
NVIDIA RAPIDS libraries for data science
NVIDIA highlights workflows that can migrate to DGX Cloud or to larger NVIDIA-accelerated data centers after local prototyping.[3][1]
Independent reviewers (LMSYS, Simon Willison) noted that while support was rapidly improving around launch, the broader ARM64 developer ecosystem (for example wheels for certain framework versions) was still maturing, and NVIDIA's official containers and playbooks eased setup for early adopters.[11][12]
NVIDIA positions Spark primarily for:
Prototyping AI models and applications locally
Fine-tuning models up to ~70B parameters
Inference and evaluation with models up to ~200B parameters in unified memory
Edge computing solutions for robotics (NVIDIA Isaac), smart cities (NVIDIA Metropolis), and more[15]
Agentic AI development: Creating and testing complex AI agents and physical AI applications
Computer vision development, as demonstrated by early evaluator Roboflow[16]
Data science and machine learning workflows
| Category | Detail |
|---|---|
| Architecture | Grace Blackwell (GB10 Superchip) |
| Process Node | TSMC 3nm, 208 billion transistors |
| CPU | 20-core Arm v9.2 CPU (10 Cortex-X925 + 10 Cortex-A725) |
| CPU Cache | 32 MB L3 cache (16 MB per cluster) |
| GPU | Blackwell-generation GPU with 5th-gen Tensor Cores; 6,144 CUDA cores |
| GPU Cache | 24 MB L2 cache |
| AI Performance | Up to 1 PFLOP (theoretical) at FP4 with sparsity; 31 TFLOPS FP32 |
| Unified memory | 128 GB LPDDR5x-9400, coherent, unified; ~273 GB/s bandwidth |
| Storage | 4 TB self-encrypting NVMe M.2 (base config) |
| Networking | 1× RJ-45 10 GbE; ConnectX-7 Smart NIC (QSFP), up to 200 Gb/s |
| Wireless / BT | Wi-Fi 7, Bluetooth 5.4 |
| Display / I/O | 1× HDMI 2.1a; 4× USB-C (one used for power) |
| Power | 240 W external PSU, USB-C PD input; TDP 140W |
| Dimensions / weight | 150×150×50.5 mm; ~1.2 kg |
| OS | NVIDIA DGX OS (Ubuntu 24.04-based) |
While the "Spark" in the product's name is primarily a branding term alluding to ignition and creation, the DGX Spark system is fully capable of running accelerated data analytics workloads.
NVIDIA separately develops the RAPIDS Accelerator for Apache Spark, a software plugin that uses NVIDIA GPUs to accelerate Apache Spark pipelines with no code changes.[17][18] The DGX Spark hardware supports the NVIDIA RAPIDS libraries, enabling it to function as a powerful, compact platform for developing and testing these GPU-accelerated big data workflows.[15]
Early third-party testing characterizes DGX Spark as an excellent developer platform whose strengths are compactness, unified memory capacity, and integrated software, while noting that raw throughput can trail larger discrete-GPU workstations due to LPDDR5x bandwidth limits. LMSYS reported that Spark "shines" for smaller models and batched inference and can run very large models (for example 70B, 120B) for experimentation, but discrete GPUs outpace it on some benchmarks.[11] The system's memory bandwidth of 273 GB/s has been compared to competitors like Apple's Mac Studio, which offers approximately 400 GB/s with the M4 Max.[19]
Micro Center's hands-on similarly framed Spark as ideal for local fine-tuning and iterative development rather than a replacement for high-end training rigs.[20]
Early reviewers have praised the DGX Spark's hardware capabilities while noting that the software ecosystem is still in early development stages.[12] Broader tech coverage traced the product's evolution from Project Digits and emphasized its role in bringing "personal AI supercomputers" to deskside workflows, with partners preparing GB10-based alternatives.[2][6]
Some commentators and community posts have criticized value-for-money relative to other hardware, especially for pure throughput tasks, while acknowledging software polish and developer ergonomics as differentiators.[21]
Early evaluators of the platform included a wide range of companies and research organizations: Anaconda, Cadence, ComfyUI, Docker, Google, Hugging Face, JetBrains, LM Studio, Meta, Microsoft, Ollama, Roboflow, and the NYU Global AI Frontier Lab.[1][16]
The DGX Spark Founder's Edition was listed at US$3,999 on the NVIDIA Marketplace at launch. Channel partners and OEMs announced availability windows in October 2025.[4][13][1]
| Region | Price | Notes |
|---|---|---|
| United States | $3,999 USD | MSRP before taxes |
| United Kingdom | £3,700 GBP | Including VAT |
| Germany | €3,689 EUR | Including VAT |
| Japan | ¥899,980 JPY | Retail price |
DGX Spark is available through NVIDIA directly and authorized partners including:
Acer
ASUS
Dell Technologies
GIGABYTE
HPI
Lenovo
MSI
Micro Center (US retail)
PNY (Europe)
Several OEM partners have developed systems based on the GB10 Superchip with similar specifications:[22]
Dell Pro Max with GB10
ASUS Ascent GX10[23]
Acer Veriton GN100
NVIDIA CEO Jensen Huang personally delivered the first DGX Spark units to several prominent figures in the AI industry, including:[24]
Elon Musk at SpaceX's Starbase facility in Texas
Sam Altman of OpenAI
NVIDIA DGX Station
DGX Cloud
NVIDIA Grace Hopper Superchip
Unified memory
ConnectX-7
NVLink-C2C
Apache Spark
NVIDIA RAPIDS
AI accelerator
Edge computing