NVIDIA has recently rolled out the DGX Spark, a desktop AI computer boasting 64GB of unified memory. Priced starting at $4,999, this new offering is set to be officially launched on October 23, available through various OEM manufacturers. The latest model maintains the GB10 Grace Blackwell superchip, DGX OS operating system, a comprehensive NVIDIA AI software suite, and ConnectX-7 networking functionalities. The sole modification is a decrease in unified memory capacity, from the initial 128GB to 64GB. Despite this reduction, the DGX Spark remains capable of locally running models with up to 100 billion parameters, catering to the majority of workloads that do not necessitate extensive training fine-tuning, such as local inference and AI agent development. The decrease in memory capacity has not compromised its fundamental computing prowess for model adaptation. Moreover, this variant facilitates multi-machine collaboration, enabling two units to be directly linked via QSFP cables to construct a local AI cluster, creating a 128GB memory pool that supports running models with up to 200 billion parameters, with performance enhanced up to 1.7 times that of a single unit.

This strategic move comes in response to a surge in demand for AI servers, resulting in a constrained memory supply. Consequently, NVIDIA has decided to increase the price of the 128GB Founders Edition version to $6,950. Although the 64GB version is priced lower than the current 128GB iteration, it still reflects an uptick compared to the initial launch price of $3,999 for the first-generation 128GB model. NVIDIA is progressively repositioning the DGX Spark from a standalone desktop AI supercomputer to a scalable local AI node, aligning with the prevailing trend of model lightweighting and effectively tackling industry hurdles associated with the tight memory supply.