⌘K
← All GPUs

NVIDIA DGX Spark

NVIDIA · AI Workstation · 2025

128 GB VRAM
$3,999 MSRP

Performance Benchmarks

LLM Inference all sourced rows →

ModelMetricValueSource
DeepSeek-V4-Flash (0731) MXFP4
2x networked DGX Spark; community reports 20-30 tok/s depending on load; recorded midpoint as estimate
 tokens per second 25 tok/s est
llama.cpp (experimental) · 2026-09-01
 Unverified community www.reddit.com ↗

Full Specifications

Source url
https://www.nvidia.com/en-us/products/workstations/dgx-spark/
Source name
NVIDIA DGX Spark Official Page
Limitation
ARM-based, no Windows support, not user-upgradeable
Ai compute fp4 pflops
1
Verdict
Best for developers who need 128GB unified memory for large model prototyping without multi-GPU complexity.
Best for
Large model prototyping (up to 200B)
Rocm support
No
Configuration verified at
2026-07-23
Noise load dba
Noise idle dba
Network infiniband
Network ethernet
Pcie lane notes
Pcie x16 slots
Cpu threads
Chassis model
Release year
2025
Onsite support
No
Warranty years
3
Windows support
No
Metal support
No
Cuda support
Yes
Wsl support
No
Linux support
Yes (DGX OS)
Operating system
NVIDIA DGX OS (Ubuntu-based)
Weight kg
2.3
Cooling type
active air
Dimensions
202 × 202 × 63 mm
Estimated peak power watts
150
Storage primary type
NVMe SSD
Power supply watts
150
Storage primary gb
4096
System memory gb
128
System memory type
LPDDR5x unified
Gpu memory bandwidth each gbps
273
Combined memory architecture
unified (LPDDR5x)
Unified memory gb
128
Total gpu vram gb
128
Gpu vram each gb
128
Gpu vendor
NVIDIA
Gpu count
1
Cpu cores
20 (10 Cortex-X925 + 10 Cortex-A725)
Gpu model
NVIDIA Blackwell (integrated in GB10)
Cpu model
NVIDIA Grace Blackwell (GB10)
Cpu vendor
ARM
Form factor
compact desktop

Related GPUs

Common Questions

Can the NVIDIA DGX Spark run local LLMs?

Yes, if the model fits in its 128 GB of memory — VRAM capacity decides which models load at all, while its memory bandwidth decides how fast tokens generate. With 24 GB or more it can hold large quantized models such as 70B-class LLMs in reduced precision, and our model pages list exact VRAM fits per model.

How much VRAM does the NVIDIA DGX Spark have?

NVIDIA DGX Spark has 128 GB of video memory , per the specification values in our hardware database.

Is the NVIDIA DGX Spark worth it for AI workloads?

At a $3,999 MSRP, its value depends on whether your models fit in 128 GB. Compare its VRAM and bandwidth against your target models with our VRAM calculator, or see how it stacks up in head-to-head comparisons.

Guides & Tools

CompareAIHardware.com participates in the Amazon Associates program. As an Amazon Associate, we earn from qualifying purchases.