NVIDIA DGX Spark, Arm, 10 Cortex-X925 + 10 Cortex-A725 Arm, 128GB, 4TB SSD

★★★★☆ 4.0 (27)

Der 20-Kern-Arm-Prozessor mit 3,8 GHz und die integrierte NVIDIA-Blackwell-Architektur liefern bis zu 1 PetaFLOP KI-Leistung in einem kompakten 3,02-kg-Gehäuse. Mit 128 GB DDR5-Speicher verarbeitet er Modelle mit bis zu 200 Milliarden Parametern in FP4, und die 4 TB NVMe-SSD unterstützt lokale Entwicklung. Dieses Mini-Supercomputersystem ist für KI-Entwickler gedacht, die große Modelle direkt am Schreibtisch feinabstimmen, testen und prototypisieren möchten, ergänzend zu Cloud- oder Rechenzentrumsressourcen.

CPU ARM Cortex Cortex-A725
RAM 128 GB
Storage 4 TB
GPU NVIDIA Blackwell Architecture
Form Factor mini
OS NVIDIA DGX OS
NVIDIA DGX Spark, Arm, 10 Cortex-X925 + 10 Cortex-A725 Arm, 128GB, 4TB SSD desktop
45 Gesamtbewertung

Derzeit kein aktueller Händlerpreis

Überblick

The 30-Second Version

The DGX Spark is a desk-sized AI supercomputer with 128GB of unified memory and a 4TB SSD, both near the top of our charts. It's built for running 200B-parameter models locally, not gaming or general use. At $8,099, it's worth it for serious AI work, but everyone else should look elsewhere.

Bild herunterladen

Pros & Cons

Vorteile

  • 128GB of unified memory means huge models run locally 99th
  • 4TB NVMe storage gives you room for big datasets 98th
  • Compact 3kg design fits on a desk
  • Full NVIDIA AI software stack preloaded

Nachteile

  • GPU performance for non-AI tasks is near the bottom
  • Reliability scores are concerning at 12th percentile
  • Only 5th percentile for ports, so connectivity is limited
  • Thermal issues mentioned in early reviews

Was Besitzer sagen

The Word on the Street

4.0/5 (27 reviews)
👍 Several buyers are using it for data science programs and love how well it handles their coursework and model work.
👎 A few owners report thermal issues that show up shortly after setup, with one saying they'll rebuy once it's fixed.

Wie sich die Meinung der Besitzer im Lauf der Zeit verändert hat

Exklusiv

Basierend darauf, wann Kunden ihre Bewertungen tatsächlich geschrieben haben - so sehen Sie, ob das anfängliche Lob anhielt.

9 geschriebene Rezensionen, Nov. 2025 – Feb. 2026: 6 zufrieden, 3 unzufrieden

Nov. 2025: 1 unzufriedenDez. 2025: 1 zufrieden, 1 unzufriedenJan. 2026: 3 zufrieden, 1 unzufriedenFeb. 2026: 2 zufriedenNov. 25Dez. 25Jan. 26Feb. 26Nov. 2025: 1 unzufriedenDez. 2025: 1 zufrieden, 1 unzufriedenJan. 2026: 3 zufrieden, 1 unzufriedenFeb. 2026: 2 zufriedenNov. 25Dez. 25Jan. 26Feb. 26
Zufrieden (4-5★)Gemischt (3★)Unzufrieden (1-2★)

Über alle 27 Shop-Bewertungen: 4,0★. Die meisten Besitzer bewerten, ohne eine Rezension zu schreiben. Das Diagramm zählt die 9 geschriebenen Rezensionen.

Jeder Punkt ist eine geschriebene Rezension, im Monat, in dem sie verfasst wurde. Zufrieden = 4-5★, gemischt = 3★, unzufrieden = 1-2★.

Die Fakten

Performance

The RAM and storage are the stars here. 128GB of DDR5 lands in the 99th percentile, and the 4TB NVMe SSD sits at 98th. That unified memory is the whole point, letting you load massive models that would choke a normal workstation. The CPU is fine, middle of the pack at 52nd percentile, but the GPU ranking is rough at 8th percentile. That's because the Blackwell chip is optimized for AI workloads, not gaming or traditional graphics. For model fine-tuning and inference, it's a beast. For anything else, it's a square peg in a round hole.

Performance Percentiles

CPU 52.1
GPU 8.1
RAM 98.9
Anschlüsse 5.3
Speicher 98.1
Zuverlässigkeit 11.5
Nutzerresonanz 54.7

Specifications

Full Specifications

Processor

CPU ARM Cortex Cortex-A725
Cores 20
Frequency 3.8 GHz

Graphics

GPU NVIDIA Blackwell Architecture
Type Integrated

Memory & Storage

RAM 128 GB
RAM Generation DDR5
Storage 4 TB
Storage Type NVMe SSD

Build

Form Factor mini
Weight 3.0 kg / 6.7 lbs

System

OS NVIDIA DGX OS

vs Competition

The Mac Studio M5 Max is the obvious alternative for creative pros, but it can't touch the DGX Spark's 128GB of unified memory for AI workloads. The Lenovo Legion Tower 5a and Dell Area-51 are gaming machines first, and they'll crush the Spark in traditional GPU tasks. The CLX SET and MSI Aegis are similar stories, built for gaming and general use. None of them are trying to do what the Spark does. If you need local AI compute, the Spark is in a class of one. If you need a general-purpose desktop, all five competitors are better choices.

Spec NVIDIA DGX Spark Apple Mac Studio M5 Max Lenovo T Series Towers Legion Tower 5a Gen 10 Dell Desktop ECT1250 HP OMEN GT16-0364 ASUS ROG GM700
CPU ARM Cortex Cortex-A725 Apple M5 Max AMD Ryzen 7 7800X3D Intel Core Ultra 7 265 AMD Ryzen 7 8700F AMD Ryzen 7 9800X3D
RAM (GB) 128 36 32 32 32 32
Storage (GB) 4096 1024 4096 2048 1024 2048
GPU NVIDIA Blackwell Architecture Apple (32-Core) NVIDIA GeForce RTX 5070 Ti Intel UHD Graphics NVIDIA GeForce RTX 5070 NVIDIA GeForce RTX 5070 Ti
Form Factor mini mini mid-tower sff mid-tower mid-tower
Psu W - 480 850 180 850 850
OS NVIDIA DGX OS macOS Windows 11 Home Windows 11 Home Windows 11 Home Windows 11 Home
Compare Compare Compare Compare Compare
Produkt CPUGPURAMAnschlüsseSpeicherZuverlässigkeitNutzerresonanz
NVIDIA DGX Spark 52.18.198.95.398.111.554.7
Apple Mac Studio M5 Max Compare 95.785.292.48968.899.366.8
Lenovo T Series Towers Legion Tower 5a Gen 10 Compare 73.494.777.988.398.17099.9
Dell Desktop ECT1250 Compare 88.260.381.674.582.77098.2
HP OMEN GT16-0364 Compare 69.590.887.197.953.57098.2
ASUS ROG GM700 Compare 80.594.781.6969037.399.3

Preis

Value & Pricing

At $8,099, this is a serious investment. But compare it to renting cloud GPUs for a year of heavy model training, and the math starts to make sense for the right buyer. You're paying for the unified memory architecture and the ability to iterate locally without metered cloud costs. For a hobbyist or casual user, it's absurd. For a data science team or serious AI researcher, it could pay for itself in months.

Mehr erfahren

Overview

The DGX Spark is NVIDIA putting a datacenter on your desk. It packs a 20-core Grace CPU, 128GB of unified memory, and a Blackwell GPU into a 3kg box. The pitch is simple: run 200-billion-parameter models locally without renting cloud GPUs.

It's not a gaming PC. It's not even a normal workstation. This is a specialized tool for AI developers and data scientists who need serious local compute. At $8,099, it's expensive, but the hardware inside is genuinely unusual for a desktop form factor.

Common Questions

Q: Can I use this as a regular desktop PC?

Technically yes, but it's not built for that. The GPU ranks in the 8th percentile for traditional graphics tasks, so gaming and video editing will be underwhelming. It's designed for AI workloads.

Q: How big of a model can I run locally?

NVIDIA says you can run models up to 200 billion parameters at FP4 precision thanks to the 128GB of unified memory. That's enough for most open-source LLMs without touching the cloud.

Q: Is the thermal issue a dealbreaker?

It's worth watching. Early reviews mention thermal problems, and reliability scores are low at the 12th percentile. If you're buying for production work, keep an eye on firmware updates and cooling solutions.

Who Should Skip This

Skip this if you need a general-purpose desktop. Gamers will be miserable with an 8th percentile GPU, and creative pros will find the port selection at 5th percentile frustrating. If you're not actively working with large AI models, you're paying $8,099 for hardware you won't use properly.

Verdict

Buy this if you're training or fine-tuning large models and want to do it locally. Data scientists, AI researchers, and developers working with LLMs will get the most out of it. The 128GB of unified memory is the killer feature, and nothing else in this form factor comes close. Just know what you're signing up for: this is a specialized tool, not a do-everything desktop.

Usage Scores

Gesamt (45.2)AI/LLM (22.8)Gaming (10.1)Mobilität (39.4)Kreative (21)Business (41.7)Entwicklung (51.3)Home Office (52.6)Workstation (47.8)
  • Gesamt45.2
  • AI/LLM22.8
  • Gaming10.1
  • Mobilität39.4
  • Kreative21
  • Business41.7
  • Entwicklung51.3
  • Home Office52.6
  • Workstation47.8

Weitere Konfigurationen3

Ähnliche Produkte