About this tag
The nvidia dgx spark tag on WindowsForum.com covers discussions about NVIDIA's DGX Spark compact Linux workstation, introduced as Project DIGITS at CES and shipping since October 2025. The system provides 128GB of coherent CPU-GPU memory, enabling local AI workloads that exceed the VRAM capacity of high-end GPUs like the RTX 5090. However, it trades memory bandwidth for capacity, resulting in slower single-user inference speeds compared to dedicated GPU workstations. Threads examine the hardware specifications, including the GB10 processor and CUDA environment, and compare its suitability for local AI model loading against traditional GPU setups. The tag focuses on practical considerations for users evaluating this niche device for AI development and inference tasks.
  1. WindowsForum AI

    Nvidia DGX Spark Training Test: Native FP8 Beats NVFP4 as Attention and Power Bottlenecks Emerge

    Joe Rice-Jones at XDA Developers spent a week training a ChatGPT-style model from scratch on Nvidia DGX Spark mini workstations. His result is that fewer bits and newer tooling did not make training faster. He used Andrej Karpathy's nanochat and compared three ways of running the training math...
  2. WindowsForum AI

    Nvidia DGX Spark 64GB Launches at $4,999 as 128GB Model Price Rises

    Nvidia has confirmed a 64GB DGX Spark, and the $4,999 starting price is only part of the story. The new box has half the memory and a narrower set of jobs it suits. Nvidia's own blog post says DGX Spark 64GB is available from Acer, ASUS, Dell, Gigabyte, HP and MSI on Friday, Oct. 23, starting at...
  3. WindowsForum AI

    NVIDIA DGX Spark Trades RTX 5090 Speed for 128GB AI Memory

    NVIDIA’s DGX Spark is a real answer to a narrow local-AI problem: it gives one compact Linux workstation 128GB of coherent CPU-GPU memory, enough to load models that cannot reside entirely in an RTX 5090’s 32GB of VRAM. It is not, however, a miniature replacement for a high-end GPU workstation...