Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents: 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
ReleaseCode / SDK · ActionableHabr AI→ original

Nvidia Builds a 512 GB Mini-Cluster from Four DGX Spark Units for Models Up to 700B Parameters

When you run large models at your desk, the bottleneck is memory, not speed: the weights simply don't fit. At SC26, Nvidia showed the DGX Spark HPC Developer Kit, four mini-computers with 128 GB each, for 512 GB of memory in total. The company says this is enough to run models up to 700B parameters locally, without the cloud. Nvidia has not disclosed a price, but four DGX Spark units bought separately cost almost $28,000.

AI-processed from Habr AI; edited by Hamidun News
Nvidia Builds a 512 GB Mini-Cluster from Four DGX Spark Units for Models Up to 700B Parameters
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Each node20 Arm cores, a Blackwell GPU, up to 1 petaflop in FP4
Key metric #2Nodes are linked in a ring via ConnectX-7, so no switch is needed
Key metric #316 TB of NVMe in total; HPC SDK with C, C++ and Fortran compilers

The main problem with running large language models locally is not compute speed but memory: if the weights don't fit, the model simply won't start. At the SC26 conference in Chicago, Nvidia showed the DGX Spark HPC Developer Kit, a cluster of four compact DGX Spark computers. Each has 128 GB of memory, for 512 GB in total.

According to the company, that is enough to run models of up to 700B parameters locally. The simplest way to picture it: a 700B-parameter model in a 4-bit format takes roughly 350 GB (this is our own estimate, not an Nvidia figure). It won't fit on a single machine with 128 GB, but it will fit on four machines combined.

For a developer, this means they can work with proprietary code or data on their own premises, without renting cloud GPUs.

What's inside

Each node is built around the GB10 Grace Blackwell superchip: a 20-core Arm processor and a Blackwell GPU. A node has 128 GB of LPDDR5X memory shared between the CPU and GPU, so data does not need to be copied between system RAM and video memory. Peak performance of a single node is 1 petaflop in FP4 operations. A petaflop is a quadrillion operations per second, and FP4 is a 4-bit number format often used to run compressed models. Across the four nodes, that adds up to about 4 petaflops on the same metric. Total NVMe storage is 16 TB.

The nodes are linked in a ring through Nvidia ConnectX-7 network cards: each DGX Spark connects to two neighbors. As a result, no separate network switch is needed, which simplifies assembly on a desk or in a small lab. The announcement gives no details on the data transfer speed between nodes.

Who it's for and what about software

The kit is aimed at researchers and developers who need to run AI models locally, debug distributed applications and carry out resource-intensive calculations. Imagine a team writing a scientific program for a large supercomputer. They can first test the code on four nodes in their own lab, then submit it to the queue of an expensive system.

For such tasks, Nvidia is preparing deployment instructions for the HPC SDK. The package includes GPU-accelerated C, C++ and Fortran compilers, libraries and MPI message-passing tools for scientific applications.

Price and what to watch for

The kit can already be ordered, but Nvidia does not list a price on its website: for details, you have to submit a request. Retail prices offer a benchmark: four separate DGX Spark units with 128 GB of memory and a 4 TB SSD cost almost $28,000. That excludes networking and additional hardware. How much the ready-made kit will cost is unknown.

The practical takeaway is this. If the workload is constant and the data is sensitive (internal code, medical or legal documents), a local cluster can be a reasonable alternative to the cloud. If you need the model only occasionally, renting will most likely stay cheaper, but you should calculate your own workload to reach a firm conclusion.

There is also a caveat. The 700B-parameter claim comes from the manufacturer, and there are no independent measurements yet of generation speed on models that large. LPDDR5X memory is typically slower than the memory in server accelerators, so response speed on the largest models may turn out to be modest. Before buying, it makes sense to wait for the first independent tests and compare them with your requirements.

Story

Key Questions Answered

What's inside?

Each node is built around the GB10 Grace Blackwell superchip: a 20-core Arm processor and a Blackwell GPU. A node has 128 GB of LPDDR5X memory shared between the CPU and GPU, so data does not need to be copied between system RAM and video memory. Peak performance of a single node is 1 petaflop in FP4 operations. A petaflop is a quadrillion operations per second, and FP4 is a 4-bit number format often used to run compressed models. Across the four nodes, that adds up to about 4 petaflops on the same metric.

Who it's for and what about software?

The kit is aimed at researchers and developers who need to run AI models locally, debug distributed applications and carry out resource-intensive calculations. Imagine a team writing a scientific program for a large supercomputer. They can first test the code on four nodes in their own lab, then submit it to the queue of an expensive system. For such tasks, Nvidia is preparing deployment instructions for the HPC SDK.

What are the key details on Price and what to watch for?

The kit can already be ordered, but Nvidia does not list a price on its website: for details, you have to submit a request. Retail prices offer a benchmark: four separate DGX Spark units with 128 GB of memory and a 4 TB SSD cost almost $28,000. That excludes networking and additional hardware. How much the ready-made kit will cost is unknown. The practical takeaway is this.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…