26.02.28 - [Review] NVIDIA DGX Spark: Super AI Power, Compact and Local

NVIDIA DGX Spark: The Power of a Super AI, Compact and Local

This time, Ask Corporation kindly lent us an MSI NVIDIA DGX Spark, and we ran it through various tests with our AI generation tools.

■ What Is the NVIDIA DGX Spark?

Source: https://www.nvidia.com/ja-jp/products/workstations/dgx-spark/

The NVIDIA DGX Spark is a compact desktop AI supercomputer officially released by NVIDIA in 2025. Equipped with the "GB10 Superchip," built on the Grace Blackwell architecture, it packs data-center-class AI computing power into a body measuring just 150mm × 150mm × 50.5mm.

🔑 Key point: The NVIDIA DGX Spark is an "AI supercomputer that fits on your desk," bringing AI processing power once achievable only in massive data centers directly into individual and corporate offices.

Thanks to its memory architecture featuring 128 GB of unified memory (shared between the GPU and CPU), it can run local inference on AI models with up to 200 billion parameters. By connecting two NVIDIA DGX Sparks via ConnectX-7 networking, it can even handle models with up to 405 billion parameters.

Key Use Cases

• Prototyping, testing, and validating AI models
• Fine-tuning models with up to 70 billion parameters
• AI generation and processing of high-resolution images and video (ComfyUI, etc.)
• Edge AI application development (Isaac, Metropolis, Holoscan frameworks)
• Data science and machine learning workflows

※ MetAI Real-World Use Case: On-Site Demos at Client Offices

In our day-to-day operations (MetAI LLC), one of the biggest advantages of the NVIDIA DGX Spark is the ability to bring it directly to client offices for demonstrations. We frequently receive requests from clients who want to try out MetAI's AI tools in person at their own offices, but traditional desktop workstations are too large and heavy to transport practically.

The NVIDIA DGX Spark, weighing just 1.2kg with a compact 150mm cube body, can be carried in a bag and brought directly to a client's location. Setup on-site takes only a short time, allowing demos and tests to begin right away—letting clients experience the AI's real-world capabilities firsthand during meetings and proposals. This has been extremely well received.

■ Test Environment

For this test, we evaluated processing speed and resource usage of an AI creative workflow using ComfyUI on our in-house MSI NVIDIA DGX Spark setup.

■ ComfyUI Installation Method

Installing ComfyUI on the MSI NVIDIA DGX Spark requires special steps for the Blackwell architecture, differing from a typical Linux environment. Since no solution was available on the official site or support channels, we established the following method with AI's help. We hope this is useful to others facing the same challenge.

0. Open a Command Prompt

‍Navigate to the pre-installed ComfyUI folder in the command prompt

1. Check System

nvidia-smi
nvcc --version
‍

2. Create a Python Virtual Environment (Python 3.12+ recommended)

python3 -m venv comfy_blackwell
source comfy_blackwell/bin/activate
pip install --upgrade pip setuptools wheel

3. Install PyTorch (Blackwell-Optimized Build)

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

4. Install ComfyUI and Dependencies

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
pip install -r requirements.txt

5. Launch with GB10 Optimization Options

Specifying the --highvram and --fp8_e4m3fn-text-enc flags is essential for the NVIDIA DGX Spark:

python main.py --listen 0.0.0.0 --highvram --fp8_e4m3fn-text-enc

💡 Once launched, install models (checkpoints, LoRAs, etc.) and workflows as usual. Using ComfyUI Manager makes this process more efficient.

■ Starting the Tests

All benchmarks below were performed within ComfyUI. The "first run" reflects a cold start (including initial model loading), while the "second run" reflects execution time after warm-up.

📊 Important: Because the DGX Spark uses unified memory (shared between CPU and GPU), VRAM usage always displays as 0%. Instead, processing status is monitored via RAM and GPU utilization.

① Lighting Migration (Lighting Change)

After warm-up, processing stabilized at 43 seconds, with GPU utilization at 88% and a temperature of 51°C—confirming stable operation. Notably, the temperature was about 20°C lower than comparable setups, demonstrating quiet, power-efficient performance.

Backlight

Multi-Light

Lighting from the Left

② 20x Upscale (832×1248 → 16,640×24,960)

Even for this ultra-high-resolution 20x upscale operation (final resolution: 16,640 × 24,960px), warm execution took only about 85 seconds. Thanks to the 128GB of unified memory, the process completed without any VRAM-related out-of-memory (OOM) errors.

③ Image Realism Enhancement (Img2Img High-Quality Rendering)

④ Pose and Camera Angle Change

‍⑤ ControlNet Pose Change

■ Video AI Generation Tests

① 480p → 4K Upscale (Video)

Video link:https://drive.google.com/file/d/1Zq-hvuYf14rDSDEDM_riWyubBWZiYvvy/view?usp=sharing

4K upscaling of video involves a massive amount of processing and takes time, but thanks to 128GB of memory, even long-form video can be processed without crashing due to VRAM shortages—making this highly practical for broadcast and video production use cases.

② Black & White → Color (17-Second Video)

Video link:https://drive.google.com/file/d/1Rb7h_-6546kpukRC16HlQ3hqbyNZ1BL1/view?usp=sharing

Colorization of a 17-second video was completed in about 40 seconds—a practical speed for applications such as broadcast footage restoration and automatic colorization for digital archiving.

■ Overall Assessment
The MSI NVIDIA DGX Spark brings new possibilities to AI creative workflows by offering "data-center-class performance in a compact form factor." While cold starts can be somewhat slow, warm execution delivers stable, high-speed processing. Most importantly, the absence of VRAM constraints makes it possible to handle ultra-high-resolution images and long-form video—tasks that were previously impossible with conventional GPUs.

Advantages of the NVIDIA DGX Spark

• 128 GB of unified memory eliminates VRAM shortages (OOM errors)
• Low operating temperature (50–65°C) and very quiet operation (35dB), ideal for long-duration use
• Compact, power-efficient design (1.2kg, 240W) makes office deployment easy
• Fits in a bag for client visits, enabling AI demos to start immediately on-site
• Scales out to support models with up to 405 billion parameters when connecting two units
• Comes pre-installed with NVIDIA's complete AI software stack

‍Points to Note

• Long cold-start times require a preloading strategy when running multiple models
• Setting up ComfyUI for Blackwell is complex and requires a CUDA 13.0–specific build
• Video AI generation involves heavy processing loads, with longer videos taking proportionally more time

MetAI LLC will continue testing AI creative workflows using the NVIDIA DGX Spark, driving forward the implementation of AI solutions for broadcasters and businesses alike.