With a format reminiscent of a mini PC, Nvidia has presented DGX Spark, a team that the brand defines as its first personal supercomputer for artificial intelligence. Its objective is clear: to bring to the developer's table a power that previously required data centers or the cloud.
The proposal stands out for allowing inferences and local model adjustments, with unusual figures for a desktop plugged into a conventional outlet. According to Nvidia, in reduced precision for AI (FP4) it offers up to 1 petaFLOP of performance, sufficient to serve models with 200.000 billion parameters without leaving home, the laboratory or the office.
What is DGX Spark and who is it for?
DGX Spark is not a consumer PC or a gaming rig; it is a development tool. It works with DGX OS (a variant of Ubuntu with Nvidia's AI software pre-installed) and is geared towards researchers, developers and students who need to experiment, refine and validate projects in their own environment.
The philosophy of use is simple: from your usual equipment you use traditional apps and, when AI is needed, you delegate the load to the DGX Spark through the local network. This way you avoid cloud dependencies, latencies and variable costs, maintaining control over data and models.
Architecture and key specifications
The heart of the team is the GB10 Grace Blackwell, a superchip that combines CPU and GPU under the same roof to maximize the memory coherence and reduce bottlenecks. The CPU incorporates 20 Arm cores (10 high-performance Cortex-X925 and 10 efficient Cortex-A725), with 128GB of LPDDR5X unified memory which prevents data transfers between RAM and VRAM.
In storage, the system supports a NVMe M.2 SSD up to 4TB with self-encryption, and connectivity includes 7 WiFi, Bluetooth 5.4, a row 10GbE, HDMI y four USB-C. The maximum power is around 240 W, which helps maintain the compact format and contained noise, which is a bit tricky on stations with high-end discrete GPUs.
The platform is ready for high-speed links and network capabilities such as ConnectX-7, designed for advanced development scenarios. In practice, the chassis can be placed on any desktop and, with a weight of around 1,2 kg, it is easy to transport between rooms or venues.
Local AI performance
Working in FP4, DGX Spark achieves approximately 1 petaFLOP, a figure that places it halfway between a compact workstation and a small laboratory server. With this ceiling, it is capable of serving locally LLM up to 200B of parameters, something that until now was prohibited in the cloud or much larger equipment.
For projects larger than that size, Nvidia notes that it is possible connect two Spark systems and tackle even larger models, with references of up to 405.000 billion parameters. In addition, the company ensures compatibility with models of DeepSeek, Meta, NVIDIA, Google, Qwen and in general with open source alternatives that can adapt to the environment.
Ecosystem and manufacturer versions
Beyond the official Nvidia model, the platform will give rise to equipment from brands such as Acer, Asus, Dell, Gigabyte, HP, Lenovo and MSI, which will be launched their own versions with this same foundation. The company has also announced the existence of a larger desktop system for those who need more memory and expansion, and maintains in its catalog options such as DGX Station for larger-scale scenarios.
It is worth insisting: It's not a Windows PC nor is it game-oriented. The value is in prototype, refine, and infer with local AI, and being able to take that work to the cloud or data center when the project grows.
Price and Availability
Nvidia confirms a price of $3.999 (about 3.451 Euros before taxes) and availability from October 15 on their website. It won't be cheap for many pockets, but for certain professional profiles it may be worth it. savings on subscriptions and an improvement in privacy and response times.
First units and initial adoption
The device has already begun circulating among partners and industry experts. Jensen Huang personally delivered a unit to Elon Musk at the SpaceX facilities, a gesture reminiscent of the delivery of the first DGX-1 in 2016 to OpenAI. In parallel, organizations such as Anaconda, Cadence, Google, Hugging Face, LM Studio, Meta o Microsoft They are testing the system and adapting models and tools for local deployment.
DGX Spark about data center power to a desktop format, with a mature software ecosystem and hardware partners behind it. For those who must work with large models, the possibility of operating locally without cloud dependencies can make a difference in costs, control and agility.
