For the OS, the Spark is an ARM-based system that runs Nvidia’s DGX OS, an Ubuntu Linux-based working system constructed particularly for GPU processing. It comes with Nvidia’s AI software program stack preinstalled, together with CUDA libraries and the corporate’s NIM microservices.
Costs for the DGX Spark begin at US $3,999. Which will appear to be lots, however given the price of high-end GPUs with ample video RAM just like the RTX Professional 6000 (about $9,000) or AI server GPUs (like $25,000 for a base-level H100), the DGX Spark could characterize a far cheaper choice total, although it isn’t almost as highly effective.
The truth is, in response to The Register, the GPU computing efficiency of the GB10 chip is roughly equal to an RTX 5070. Nonetheless, the 5070 is restricted to 12GB of video reminiscence, which limits the scale of AI fashions that may be run on such a system. With 128GB of unified reminiscence, the DGX Spark can run far bigger fashions, albeit at a slower pace than, say, an RTX 5090 (which usually ships with 24 GB of RAM). For instance, to run the 120 billion-parameter bigger model of OpenAI’s current gpt-oss language mannequin, you’d want about 80GB of reminiscence, which is way over you will get in a shopper GPU.
A callback to 2016
Nvidia founder and CEO Jensen Huang marked the event of the DGX Spark launch by personally delivering one of many first models to Elon Musk at SpaceX’s Starbase facility in Texas, echoing a related supply Huang made to Musk at OpenAI in 2016.
“In 2016, we constructed DGX-1 to offer AI researchers their very own supercomputer. I hand-delivered the primary system to Elon at a small startup known as OpenAI, and from it got here ChatGPT,” Huang mentioned in an announcement. “DGX-1 launched the period of AI supercomputers and unlocked the scaling legal guidelines that drive trendy AI. With DGX Spark, we return to that mission.”
