Nvidia introduces performance improvements for DGX Spark in a new update.
Nvidia announced a software update for DGX Spark, the compact AI workstation that was shown at CES 2026 late last year under the codename Project Digits. According to Nvidia, the update delivers an average performance improvement of up to 2.5 times and expands the platform with access to the full AI Enterprise software stack.
Focus on local AI development
The DGX Spark is intended as a local development platform for AI prototyping, GPU-accelerated software development, and the local training and running of AI models. Although the computing power is comparable to that of an RTX 5070, the system is distinguished by its 128 GB of integrated memory, which can be fully allocated to the GPU.
Performance gains especially in compute-intensive tasks
The performance improvements are mainly in the compute-intensive parts of AI workflows, writes The Register. Nvidia says that token generation in LLMs remains limited by bandwidth, but that the update mainly accelerates the prefill. As a result, less time is needed between entering a prompt and generating output.
The updates include improvements for TensorRT-LLM, PyTorch, Llama.cpp and Nvidia’s own inference engine, among others. This should not only improve LLM inference, but also workloads such as fine-tuning and image and video generation.
AI Enterprise available as a subscription
Nvidia is bringing its AI Enterprise suite to the DGX Spark later this month. It contains frameworks, models and microservices for the development and deployment of AI applications. Normally, the package costs $4,500 per GPU per year, but Nvidia is exploring customized prices for Spark users. The suite remains free for developers, but a paid subscription is required to use applications and services in a production environment.