NVIDIA DGX Spark: stated specifications, scaling, and price of the local AI system

In the materials reviewed, NVIDIA DGX Spark is described as a desktop computing system for local deployment, testing, and fine-tuning of artificial intelligence models. According to the concept presented in these materials, the platform is built on the NVIDIA GB10 Grace Blackwell superchip and combines central and graphics processor resources in unified memory. This scenario is aimed at developers, researchers, and companies that need to work with large models without constantly transferring data to the cloud.

Before purchasing, it is important to match the system configuration to the model and workload for which it is intended.

What tasks is NVIDIA DGX Spark designed for

In NVIDIA's materials, NVIDIA DGX Spark is presented as a system for developing and testing AI applications directly at the workstation. NVIDIA associates it with autonomous agents, generative models, data analysis, robotics, computer vision, and edge use cases. According to the manufacturer's description, the preinstalled software stack shortens the path from hardware installation to the first experiment with a model.

• Prototyping. In the stated use cases, a developer creates and tests models, agent applications, and services that generate text or code.
• Local inference. A model responds to requests without constantly contacting a cloud server. This option may be convenient when working with sensitive datasets, provided that the selected software environment supports the required model.
• Fine-tuning. In the description of the 128-gigabyte configuration, NVIDIA states that it can fine-tune models with up to 70 billion parameters, but the specific result depends on the training method and settings.
• Science and data analysis. According to the manufacturer's concept, unified memory can simplify the processing of large datasets when data and computations need to exchange information.
• Edge development. In NVIDIA's materials, the platform is considered for preparing applications in robotics, computer vision, and other systems operating near the data source.

In the review author's assessment, DGX Spark should not be viewed as a replacement for a general-purpose server cluster in every scenario. Limitations are likely with large training batches, lengthy experiments, simultaneous work by multiple teams, and tasks requiring a large amount of disk or RAM. For such projects, a desktop system may serve as a development and preliminary testing environment, after which the workload can be moved to a data center or cloud infrastructure if a specific test confirms this is appropriate.

NVIDIA DGX Spark memory and specifications

In the description reviewed, unified system memory accessible to both the central and graphics processors is identified as the key element of DGX Spark. For the new configuration, 64 GB of LPDDR5X memory is specified. The lineup also includes a 128 GB version. The same materials state a 256-bit interface and bandwidth of 273 GB/s.

The 1 PFLOP figure refers to FP4 computations, so it cannot be directly compared with the performance of a graphics card in FP16, FP8, or FP32. Buyers should check the actual speed on the required model and under the chosen workload scenario rather than treating a single stated figure as a universal measure.

Which models fit on a single device

In the materials reviewed, the DGX Spark configuration with 64 GB of memory is given as suitable for local deployment of models with approximately up to 100 billion parameters. This is an inference estimate, not a promise of the same speed for every format and task.

For the 128 GB version, NVIDIA gives an estimate of approximately up to 200 billion parameters. It is also stated that the 128-gigabyte system can be used for inference with models of this scale and for fine-tuning models with up to 70 billion parameters.

The specific names of compatible models are not provided in the materials reviewed. Support for an individual model depends on the weight format, the available optimized runtime, and the version of NVIDIA's software stack.

How to scale NVIDIA DGX Spark

According to the NVIDIA materials reviewed, multiple DGX Spark systems can be connected into a local cluster through a high-speed NVIDIA ConnectX network. They state that up to four systems can be connected. According to the manufacturer's concept, this option increases the available unified memory and makes it possible to work with models that do not fit in a single chassis.

For combinations of multiple devices, the NVIDIA materials reviewed do not provide performance figures, bandwidth figures, or comparative test results. They cannot be calculated by simply multiplying the specifications of a single computer, since the source does not specify the resulting figures for such configurations.

Price of the new and original version

The NVIDIA materials reviewed do not specify the price of DGX Spark configurations. Therefore, the cost of the 64 GB and 128 GB versions cannot be compared based on the initial description.

Such systems cannot be compared solely by the cost per gigabyte. The final price may depend on the supply channel, package, support terms, availability from partners, and software configuration. Before ordering, it is necessary to check exactly what memory is installed, who is responsible for the warranty, whether the required network accessories are included, and whether the offer applies to the buyer's region.

When choosing between the versions, 64 GB corresponds to NVIDIA's stated guidance for running models with approximately up to 100 billion parameters. For fine-tuning and working with a larger context, the 128 GB configuration or a cluster of several systems appears more reasonable if that version is available and matches the task. This is the review author's assessment, not the result of a comparative test.

Availability and supply partners

For the 64-gigabyte configuration, the NVIDIA page reviewed lists the status as “Coming Soon” and offers the option to subscribe for a notification. This status does not confirm that the system is actually available for ordering in a specific region.

Among the manufacturers of systems based on NVIDIA GB10 named in the materials reviewed are Acer, ASUS, Dell Technologies, Gigabyte, HP, Lenovo, and MSI. The same materials mention retail and distribution partners, including Amazon, Arrow, Micro Center, PNY, TD SYNNEX, and Ingram Micro.

Delivery times depend on the country, partner, and stock levels. A company planning a local cluster should clarify in advance whether it is possible to order several identical systems: mixing different memory configurations will complicate the calculation of available capacity and model distribution.

Overall assessment of NVIDIA DGX Spark

In the review author's assessment, DGX Spark may be of interest to those who need a compact local platform for developing agents, testing large language models, analyzing data, and preparing solutions for transfer to a data center. In NVIDIA's materials, the system is described in terms of unified memory, the manufacturer's software stack, and the ability to connect multiple devices through a high-speed network.

Buyers should start not with the stated PFLOP figure but with a task matrix: model, quantization format, maximum context, need for fine-tuning, number of concurrent users, and acceptable latency. They can then compare the 64-gigabyte and 128-gigabyte versions and assess a cluster. This sequence helps determine whether the system's compactness is justified and whether there will be enough memory after the required software environment is installed.

Frequently asked questions about NVIDIA DGX Spark

Can a model with 100 billion parameters be run on DGX Spark?

For the 64 GB version, NVIDIA states an estimate of approximately up to 100 billion parameters. Actual feasibility depends on the available configuration, selected model, weight format, and software environment.

Is DGX Spark suitable for fine-tuning?

The platform description includes model fine-tuning. For the 128-gigabyte configuration, NVIDIA specifies models with up to 70 billion parameters; however, the specific training method may require a smaller model or additional optimizations.

Will a cluster of two devices make work exactly twice as fast?

The NVIDIA materials reviewed contain no test result for a combination of two devices. Simply doubling the figures for one computer cannot be considered confirmed: the source does not specify performance for this configuration.

Which should I choose: 64 or 128 GB?

64 GB corresponds to the stated guidance for models with approximately up to 100 billion parameters. 128 GB is intended, according to NVIDIA, for running models with approximately up to 200 billion parameters and fine-tuning models with up to 70 billion parameters, if that version is available and matches the task.