Summary of overseas reviews of 'NVIDIA's world's smallest supercomputer,' DGX Spark



On October 15, 2025, NVIDIA released the DGX Spark , which claims to be the 'world's smallest supercomputer.' Overseas media outlets who have actually used the DGX Spark have already published reviews, so we've compiled information on their actual usage experience.

A Grace Blackwell AI supercomputer on your desk | NVIDIA DGX Spark

https://www.nvidia.com/en-us/products/workstations/dgx-spark/

NVIDIA's DGX Spark is a compact, palm-sized supercomputer equipped with an NVIDIA GB10 Grace Blackwell Superchip with a maximum 1P FLOPS , 128GB of integrated memory, and an NVIDIA ConnectX-8 SuperNIC with a single port that achieves 800Gbps data transfer speeds. It is priced at $3,999 (approximately 600,000 yen).

NVIDIA announces desktop PC-sized AI supercomputers 'DGX Spark' and 'DGX Station,' equipped with Blackwell-generation AI-specialized GPUs for local AI development in a personal environment - GIGAZINE



To celebrate the launch of DGX Spark, NVIDIA CEO Jensen Huang delivered the first DGX Spark unit to Elon Musk, who claimed that DGX Spark delivers roughly 100 times more compute power per watt than the DGX-1, the first-ever AI computer Huang built for OpenAI in 2016.




The Register has published a review article titled 'NVIDIA's world's smallest supercomputer, DGX Spark, can process large language models (LLMs) at stable speeds,' and while it doesn't have the world's fastest GPU, it says, 'Everything works well enough.'

DGX Spark Nvidia's desktop supercomputer: first look • The Register
https://www.theregister.com/2025/10/14/dgx_spark_review/



The Register notes that DGX Spark is inferior to NVIDIA's latest GPU, the RTX 5090 , in areas such as LLM inference, fine-tuning, and image generation. However, it explains that DGX Spark can comfortably run AI models that are simply not possible with consumer graphics cards currently on the market.

In local AI development, no matter how powerful the CPU or GPU are, or how wide the memory bandwidth, AI models cannot run properly without sufficient VRAM. To achieve this, DGX Spark is equipped with 128GB of unified memory, the largest amount of VRAM available in an NVIDIA workstation GPU. However, this 128GB unified memory uses LPDDR5X , which is significantly slower than the GDDR7 used in NVIDIA's RTX 50 series.

AI workloads, such as running inference on AI models with up to 200 billion parameters, typically require multiple high-end GPUs, which can cost hundreds of thousands of yen. However, NVIDIA sacrificed some performance and bandwidth to secure pure capacity, 'building a system (DGX Spark) that's not the fastest at any particular task, but can do anything,' The Register explained.

The DGX Spark measures just 150mm wide x 150mm deep x 50.5mm high and uses a standard flow-through design, drawing air in through a metal mesh panel in the front and expelling it out the back, allowing all I/O to be located at the rear.



The NVIDIA GB10 Grace Blackwell Superchip

SoC is manufactured using TSMC's 3nm process.

The CPU does not use the ARM Neoverse used in NVIDIA's higher-end models, but instead features ten ARM Cortex-X925 high-performance cores and ten ARM Cortex-A725 high-efficiency cores.

The GPU uses the same Blackwell architecture as NVIDIA's RTX 50 series, with a performance of 1 PFLOPS at FP4 precision.

The OS is NVIDIA's DGX OS, a custom version of Ubuntu 24.04.



DGX Spark is a computer specialized for workloads such as machine learning, generative AI, and data science. While these fields are less esoteric than they once were, they can still be difficult for beginners to understand, The Register points out. However, the DGX Spark software ecosystem offers a wealth of documentation, tutorials, and demonstration videos to help you get started quickly with workloads in these fields.

The Register pointed out that 'DGX Sparks' true competitors are not consumer or workstation GPUs, but rather systems such as Apple's M4-equipped Mac Mini or Mac Studio, or AMD's Ryzen Al Max+ 395- equipped systems.' 'If you're looking for a small, low-power AI development platform that can handle everything from productivity to content creation and game development, DGX Sparks probably isn't for you. It would be wise to invest in AMD's Strix Halo or Mac Studio, or wait until Windows comes with the NVIDIA GB10 Grace Blackwell Superchip. However, if you're primarily focused on machine learning and are looking for a relatively affordable AI workstation, there are few devices that meet as many criteria as DGX Sparks.'

Data scientist Bojan Tunguz wrote, 'DGX Sparks is the dream of data science, machine learning, and AI developers everywhere. It's a small computer that can rekindle the local development fires that this community has been hoping for, and it takes full advantage of NVIDIA's hardware and software benefits in a very attractive form factor. For many years, I had to carry a huge Dell workstation with an RTX 5000 in addition to my main laptop for work. But now I can finally leave that heavy machine at home and take this small supercomputer with me.'

Honey, I've Shrunk the Supercomputer! - by Bojan Tunguz
https://bojan.substack.com/p/honey-ive-shrunk-the-supercomputer



Datasette developer Simon Willison describes DGX Spark as 'great hardware and early stages of an ecosystem,' but adds that 'it's still too early to confidently recommend this machine. My own lack of experience with CUDA, ARM64, and Ubuntu GPU machines in general meant I struggled to determine how best to use it.'

NVIDIA DGX Spark: great hardware, early days for the ecosystem
https://simonwillison.net/2025/Oct/14/nvidia-dgx-spark/



LMSYS Org, which promotes the development of open and scalable LLM, points out that 'DGX Spark has demonstrated excellent engineering for its size and power envelope, but naturally, its raw performance is limited compared to full-size discrete GPU systems.' While it can perform adequately for small AI models, it points out that for large AI models, it performs poorly compared to the NVIDIA RTX PRO 6000 Blackwell Workstation Edition and GeForce RTX 5090. Therefore, LMSYS Org writes, 'DGX Spark was not designed to compete head-on with full-size Blackwell GPUs or Ada-Lovelace GPUs, but rather to incorporate the DGX experience into a compact, developer-friendly form factor. As such, it is ideal for applications such as 'AI model prototyping and experimentation,' 'lightweight in-device inference,' and 'memory coherent GPU architecture research.''

NVIDIA DGX Spark In-Depth Review: A New Standard for Local AI Inference | LMSYS Org
https://lmsys.org/blog/2025-10-13-nvidia-dgx-spark/



Kyunghyun Cho, a scientist at New York University, wrote in an excited post, 'I was dying to get my hands on a DGX Spark on launch day. It's a small, powerful machine that fits on my desk and can run the latest open-source AI models of almost any size. Thanks to NVIDIA, my dream came true a few weeks ago, and I'm excited to see it sitting on my desk at NYU's Global AI Frontier Lab!'




They also successfully ran ICARE on DGX Spark, which Cho wrote 'shows the potential for putting a supercomputer on a clinician's desk.'




Others have used the latest paper to locally test LLM's causal understanding, writing, 'We were able to quickly and effectively evaluate the open-weight model locally ourselves.'




Lily Liu reported, 'We successfully launched gpt-oss-20B using virtual LLM on DGX Spark.' 'While this is not a rigorous benchmark, the initial test results look promising.'




'None of the YouTubers who received NVIDIA's DGX Spark have compared its performance with that of a MacBook Pro with an M4 Max, even though they are roughly the same price. The reason is that the MacBook Pro has a bandwidth of 546GB/s, while the DGX Spark has a bandwidth of just 273GB/s,' one person pointed out.




Additionally, citing data comparing the number of tokens that can be generated per second when running gpt-oss-120b on a Mac equipped with DGX Spark and an M3 Max, some people pointed out, 'This is a very poor result. It cannot be explained solely by the poor performance of Ollama (a local AI platform) or the difference in memory bandwidth (273 GB/s for DGX Spark and 400 GB/s for M3 Max).'




PC retailer Micro Center has released an unboxing video of the upcoming DGX Spark.

NVIDIA DGX Spark Unboxing and First Impression | Micro Center - YouTube


Related Posts:

in AI,   Video,   Hardware, Posted by logu_ii