Specifications

Architecture: NVIDIA Grace Blackwell
GPU: Blackwell Architecture
CPU: 20‑core Arm: 10 Cortex‑X925 + 10 Cortex‑A725
CUDA Cores: Blackwell Generation
Tensor Cores: 5th Generation
RT Cores: 4th Generation
Tensor Performance: Up to 1 PFLOP FP4
System Memory: 128 GB LPDDR5x, coherent unified
Memory Interface: 256‑bit
Memory Bandwidth: 273 GB/s
Storage: 4 TB NVMe‑M.2 with self‑encryption
USB: 4× USB‑Type‑C
Ethernet: 1× RJ‑45 10 GbE
NIC: ConnectX‑7 @ 200 Gbps
Wi‑Fi: Wi‑Fi 7
Bluetooth: BT 5.4
Audio Output: HDMI multichannel audio output
Power Supply: 240 W
GB10 TDP: 140 W
Display Connectors: 1× HDMI 2.1a, up to 3× DisplayPort over USB‑C (DP Alt Mode)
NVENC: 1×
NVDEC: 1×
Operating System: NVIDIA DGX OS
Dimensions: 150 mm L × 150 mm W × 50.5 mm H
Weight: 1.2 kg

What reviewers say

Tom’s Hardware reviews the NVIDIA DGX Spark as a well‑rounded local AI platform that delivers solid performance thanks to its GB10 SoC, ample 128 GB of RAM, and full CUDA support. The review highlights its polished software ecosystem and unique connectivity options but notes that it is expensive for users who don’t need all its features and that it does not yet run Windows.
www.tomshardware.com ↗
Signal65’s review frames the NVIDIA DGX Spark as a compact, high‑memory AI supercomputer that democratizes powerful GPU‑CPU integration for individual developers. The article praises its 128 GB unified memory, balanced CPU/GPU performance, ready‑to‑use software stack, and clustering capability while noting the $3,999 price tag as a potential drawback compared to typical consumer PCs.
signal65.com ↗
Team Vast reviews the NVIDIA DGX Spark as a compact personal AI supercomputer that excels at prototyping, fine‑tuning, and running large language models locally thanks to its GB10 Grace Blackwell Superchip, 128 GB unified memory, and integrated software stack. The review notes performance gains from recent software updates but highlights limitations such as modest memory bandwidth, early reliability issues (latency, throttling, reboot), and a high price point that may not justify the investment for production‑grade inference or large‑scale training.
vast.ai ↗
The review praises the DGX Spark’s compact, all‑in‑one design and its impressive 128 GB of coherent unified memory, which lets it run very large models without VRAM transfer overhead. It notes that the system shines for smaller, batched workloads and is well suited to prototyping and experimentation rather than production inference. The main drawback highlighted is the limited LPDDR5x memory bandwidth (273 GB/s), which bottlenecks raw performance compared with full‑size discrete GPU systems.
www.lmsys.org ↗

Our take

Based on 4 reviews

All four reviews agree the NVIDIA DGX Spark’s 128 GB unified memory and compact all-in-one design make it a capable local AI platform for prototyping and experimentation. Tom’s Hardware, Signal65, Vast, and LM Sys emphasize its ready-to-use software stack and GB10 integration as key strengths for developers. Based on their tests, we conclude it is best suited for individual researchers and developers who prioritize portability and unified memory over raw throughput.

However, reviewers split on value and performance bottlenecks: Vast and LM Sys cite modest 273 GB/s memory bandwidth and early reliability issues like throttling, while Tom’s Hardware and Signal65 highlight its polished ecosystem and clustering potential. This means buyers should weigh high cost and bandwidth limits against the convenience of a compact, software-ready system for small-batch workloads rather than large-scale production.

Pros

128 GB unified memory reduces VRAM overhead
Compact all-in-one design enables portability
Ready-to-use software stack simplifies setup
GB10 integration balances CPU and GPU performance

Cons

High price limits accessibility for some users
273 GB/s memory bandwidth bottlenecks throughput
Early reliability issues include throttling and latency

Best for: Individual developers and researchers prototyping large models locally; avoid for large-scale production inference or budget-conscious buyers.