NVIDIA DGX Spark Desk-Side AI Computer
Price on request · Request a quote for the DGX Spark or call (+356) 2138 5911 · NVIDIA DGX Spark specifications
A quote request gets you the configuration we recommend, the part code and what our set-up covers.
We supply, set up and support NVIDIA DGX Spark units in Malta.
The NVIDIA DGX Spark fits on a desk: it is 150 mm square, 50.5 mm high and weighs 1.2 kg. Inside is NVIDIA’s GB10 Grace Blackwell chip, which puts the processor and the graphics processor (GPU) on one chip with 128 GB of shared memory. That memory is what lets a box this size load a large AI model. It is for IT and data teams who want to run AI on their own documents without sending them to a cloud service.
If a pilot grows into a full rollout, on-premise AI compute covers the servers, and we can plan the wider AI project with you.
What is the NVIDIA DGX Spark for?
The NVIDIA DGX Spark is the box we use for a private knowledge-assistant pilot. One team asks questions of the company’s policies, manuals or contracts, and an open AI model answers from those documents. Everything runs on the Spark, on your network.
It also suits a data or development team that wants to test open models, fine-tune one on its own data, or build an AI agent before buying production hardware. Banks, insurers and iGaming operators work under rules on where customer data may go, so a pilot on the Spark lets them try AI without sending documents to a cloud service.
How a company knowledge assistant is built and piloted is written up step by step in our case study.
Keep company documents on your own network
The NVIDIA DGX Spark runs the AI model locally, so your documents and the questions staff ask about them stay inside your building.
No cloud service in the loop. The model, the document index and the answers all sit on the Spark. Its NVMe solid-state drive is self-encrypting, so data on the drive is encrypted at rest.
Updates without an internet link. After the first set-up, NVIDIA’s software updates can be installed from a USB drive or a local repository, so the Spark can sit on a network with no internet access.
Quiet enough for an office. NVIDIA rates it at about 29 dB at the user’s position while working and about 13 dB at idle. It needs no server room, only a desk, a power socket and the office network.
How large a model can the DGX Spark run?
NVIDIA states that one DGX Spark with 128 GB of memory runs models of up to 200 billion parameters for inference (answering questions), using models at FP4 precision. For fine-tuning (teaching a model your own material) the limit is 70 billion.
One pool of memory. The processor and GPU share the 128 GB, so a model is loaded once instead of being split between system memory and a smaller graphics card. NVIDIA rates the GB10 at up to 1 PFLOP (a thousand trillion calculations a second) at FP4, a compact 4-bit number format for AI, with sparsity (skipping zero values) switched on.
Two units for a larger model. Two Sparks linked by one cable on their 200 Gb/s QSFP ports (the fast network sockets on the back) work as one memory pool. NVIDIA states models of up to 400 billion parameters on a pair.
Ready-made software. It ships with NVIDIA DGX OS, a Linux system based on Ubuntu, and NVIDIA’s AI software preinstalled, so set-up time goes into your documents rather than drivers. Source: NVIDIA DGX Spark user guide.
How fast does the DGX Spark answer?
The NVIDIA DGX Spark reads a question quickly but writes the answer at a pace set by its memory speed, 273 GB/s. In the LMSYS team’s October 2025 tests, on early software, a 20-billion-parameter model wrote about 50 tokens (word pieces) per second, which reads comfortably. A 70-billion-parameter model wrote under 3 per second for a single user, and LMSYS rates such large models as suited to prototyping rather than production.
So we size the pilot around a model that answers at reading speed, usually a small or mid-sized open model. Larger models on the Spark are for testing, not for a team’s daily work. When the pilot proves itself and many people will use the assistant, a GPU server gives the speed.
NVIDIA DGX Spark technical specifications
| Specification | NVIDIA DGX Spark |
|---|---|
| Superchip | NVIDIA GB10 Grace Blackwell: CPU and GPU on one chip, joined by NVLink-C2C |
| GPU | NVIDIA Blackwell, 6,144 CUDA cores, 5th-generation Tensor Cores, 4th-generation RT Cores |
| CPU | 20 Arm cores: 10 Cortex-X925 and 10 Cortex-A725 |
| AI performance | Up to 1 PFLOP at FP4 with sparsity; up to 1,000 TOPS for inference |
| Memory | 128 GB unified LPDDR5x, shared by CPU and GPU; 256-bit; 273 GB/s |
| Storage | 4 TB M.2 NVMe SSD, self-encrypting (NVIDIA unit) |
| Networking | 1 x 10 GbE RJ-45; NVIDIA ConnectX-7 with 2 x QSFP ports, up to 200 Gb/s each; Wi-Fi 7; Bluetooth 5.4 |
| Ports and display | 4 x USB Type-C (one for power delivery); 1 x HDMI 2.1a with multichannel audio; up to 3 DisplayPort screens over USB-C |
| Model size | Inference up to 200 billion parameters (FP4 models); fine-tuning up to 70 billion; two linked units up to 400 billion; four units through a switch up to 700 billion |
| Operating system | NVIDIA DGX OS (based on Ubuntu Linux) with NVIDIA’s AI software preinstalled |
| Power | External 240 W power supply, included; GB10 chip rated at 140 W |
| Noise | About 29 dB sound pressure at the user’s position when working; about 13 dB at idle |
| Operating environment | 5 to 30 °C; 10 to 90% relative humidity, non-condensing |
| Size (W x D x H) | 150 x 150 x 50.5 mm |
| Weight | 1.2 kg |
See the full NVIDIA DGX Spark specifications on nvidia.com
NVIDIA DGX Spark configurations and part codes
NVIDIA’s own DGX Spark comes with 4 TB of storage. Lenovo, HP, Acer and MSI build units on the same GB10 chip with the same 128 GB of memory. We supply any of these, and a link cable if you plan two units. Tell us the task and we specify the part code. Delivery usually takes about two weeks from order.
| Configuration | Part code |
|---|---|
| NVIDIA DGX Spark Founders Edition, 128 GB unified memory, 4 TB NVMe storage | 940-54242-0005-000 |
| Lenovo ThinkStation PGX, NVIDIA GB10, 128 GB unified memory | Part code given in your quote |
| HP ZGX Nano G1n AI Station, NVIDIA GB10, 128 GB unified memory | Part code given in your quote |
| Acer Veriton GN100, NVIDIA GB10, 128 GB unified memory | Part code given in your quote |
| MSI EdgeXpert MS-C931, NVIDIA GB10, 128 GB unified memory | Part code given in your quote |
| QSFP cable for linking two units (NVIDIA-approved Amphenol or Luxshare, 0.4 or 0.5 m) | Quoted with the second unit |
The NVIDIA unit comes with its own 240 W power supply, which it needs: NVIDIA warns that another adapter can lower performance or stop it booting. We confirm the supply of a partner unit when we quote it.
How much does the NVIDIA DGX Spark cost in Malta?
The price of the NVIDIA DGX Spark in Malta is on request. It depends on the unit (NVIDIA’s own with 4 TB, or a Lenovo, HP, Acer or MSI unit on the same chip) and on what you add. These are the items we price with it:
- Storage size, set by how many documents the assistant reads.
- A second unit and a link cable, only if the model you need is larger than one unit’s memory.
- A UPS, so a power cut does not stop an update or a training run part-way.
- Software support: NVIDIA sells a separate AI Enterprise subscription for the Spark if you want vendor support for the software.
- Our set-up and pilot time: preparing the documents, choosing the model and training the team.
The whole unit runs from one 240 W adapter, and NVIDIA rates the GB10 chip at 140 W.
DGX Spark or a GPU server: which one fits?
Choose the NVIDIA DGX Spark for a pilot, a development team or one department. It needs no rack and no server room. It is the wrong choice for an assistant used by the whole company at once, or for a service that must keep running if a power supply fails: it has one external supply.
For production AI with many users, the Supermicro X13 4U Universal GPU server is the usual next step: several data-centre GPUs in a rack, sized to the real load. The Supermicro X14 5U NVIDIA GPU server is for training large models in a rack. For inference in a branch or shop with no server room, the Supermicro AS-3015TR-i4 edge tower takes one GPU in a tower. Our on-premise AI compute guide covers sizing a server for production.
If you only want a short look at a small model, a PC with a graphics card may be enough, and we will tell you so before you buy a Spark.
What we set up with an NVIDIA DGX Spark in Malta
Our sales team works out the NVIDIA DGX Spark configuration with you; our technical team installs it, connects it and hands it over. Set-up covers the first boot and NVIDIA’s software update, the office network connection and user access from staff PCs. Then comes the pilot itself: the documents, the model and a short training session for the team.
We put the Spark on a UPS. A small line-interactive unit such as the Eaton 5P Gen2 UPS covers one or two units, and we check the runtime in Eaton’s tool before quoting. More on the Eaton UPS range we supply. Keep the Spark in an air-conditioned room: NVIDIA’s ideal range tops out at 30 °C, which a closed cupboard can exceed in a Maltese summer.
Frequently asked questions about the NVIDIA DGX Spark
Do our documents leave the building with the DGX Spark?
Your documents stay in: the Spark runs the AI model and the document index locally, so questions and answers stay on your network. The first set-up downloads NVIDIA’s software over the internet; after that, updates can come from a USB drive or a local repository.
Is the DGX Spark fast enough for our staff?
The NVIDIA DGX Spark is fast enough for a pilot team with the right model. It wrote about 50 tokens per second with a 20-billion-parameter model in the October 2025 LMSYS review, and under 3 with a 70-billion model. We pick a model that answers at reading speed, and move to a GPU server when many people use it at once.
Can we start with one DGX Spark and add a second later?
The NVIDIA DGX Spark takes a second unit later: it links to the first through one QSFP cable on the 200 Gb/s ports, and NVIDIA’s software sets up the pair. NVIDIA states that two units hold models of up to 400 billion parameters, and four units connected through a network switch up to 700 billion.
Who installs it in Malta?
Our technical team installs it: the first boot and software update, the network connection, user access and the pilot set-up. Set-up needs a stable internet connection for the download, and NVIDIA warns not to switch the unit off while it updates.
Who repairs it and handles the warranty?
The NVIDIA DGX Spark comes with a one-year warranty, and we handle repairs and returns here in Malta. NVIDIA provides a built-in hardware diagnostic that checks the unit in about 30 minutes before any repair claim.
Is it held in stock in Malta?
The DGX Spark is not held in stock in Malta: we order it for you and delivery usually takes about two weeks. The Lenovo, HP, Acer and MSI units on the same chip are ordered the same way.
How much does a DGX Spark cost in Malta?
The price is on request, because it depends on the unit, the storage and what you add, such as a UPS, a second unit or our pilot set-up. Tell us the task and we quote the exact configuration.
Talk to us about a DGX Spark pilot
Tell us which documents the assistant should read, how many people will use it and where it will sit. We reply within one working day with the questions we still need answered. If the site needs a look, an engineer visits, anywhere in Malta and Gozo. You get a quote with the exact part codes, the installation scope and the warranty route.
Request a quote for the DGX Spark or call (+356) 2138 5911 · Price on request
Monday to Friday 8:30 to 16:30. We reply within one working day.
We have supplied IT and industrial equipment in Malta since 1981.
Related Products
Talk to us about your next project
Tell us what you need to run, code, label, store, protect or automate. Our technical sales team in San Gwann replies within one working day, in Maltese or English.
Who you deal with:
- Part of Maltese IT and industry since 1981
- A sales team that specifies the right part codes
- A technical team that installs, tests and repairs
- Engineers on site anywhere in Malta and Gozo
- Zebra Premier Business Partner, Synology Platinum Partner
- Distributor for Eaton and Domino in Malta
What happens next?
We reply within one working day
An engineer visits if the site needs a look
You get a quote with part codes, installation and warranty