Nvidia DGX Spark compact desktop featuring one petaflop of performance, arrives for AI model training and development

Nvidia has begun deliveries of the DGX Spark compact desktop designed for AI model development.

The small form-factor computer combines a 20-core Arm CPU (10 Cortex-X925 performance cores with 10 Cortex-A725 efficiency cores) with 128 GB of LPDDR5x RAM and 4 TB of NVMe M2 storage for up to one petaflop of Tensor core performance (FP4 using the sparsity feature) while using just 240 watts of power.

Although the system is not faster than the Nvidia RTX 5080, 5090, or Pro 6000 GPU cards based on LMSys benchmarks, its unified RAM allows AI developers to fully load large AI models into memory without the need for model quantization or GPU VRAM swapping, enabling convenient, local, offline AI model prototyping and tuning.

The SFF desktop runs Nvidia DGX OS, a customized version of Ubuntu, which is a popular Linux distribution. The computer can fit most desks due to its small 150 x 150 x 50.5 mm (5.9 x 5.9 x 2.0 in.) size and 1.2 kg (2.6 lbs.) weight.

The Nvidia DGX Spark has an MSRP of $3,999.99 and will be generally available starting October 15, 2025. The system has been listed at Microcenter, but has not yet appeared on the Nvidia store on Amazon.