NVIDIA Announces 64GB DGX Spark Desktop AI Supercomputer Starting at $4,999
NVIDIA officially introduced a 64GB unified memory version of its DGX Spark desktop AI computer starting at $4,999, powered by the GB10 Grace-Blackwell superchip. Scheduled for release on October 23, the system lowers the hardware barrier for developers building local AI agents and running large language models. As autonomous AI agent workflows consume vast numbers of tokens and drive up cloud computing costs, executing models locally offers a cost-effective and privacy-conscious alternative. The DGX Spark gives small teams and researchers desktop-class hardware to prototype, fine-tune, and deploy open-source models without depending on cloud infrastructure. The 64GB system reserves roughly 56GB of unified memory for model weights and KV cache, supporting the NVFP4 4-bit quantization format to run models like Qwen3.8-27B and Gemma4 26B smoothly. Equipped with a ConnectX-7 NIC and NVIDIA Sync tools, developers can cluster up to four DGX Spark nodes with automated setup, yielding up to 1.7x performance gains on a dual-node cluster.
## BACKGROUND
Unified memory architecture allows CPU and GPU components to share a single high-speed memory pool, overcoming traditional VRAM limitations and PCIe bandwidth bottlenecks during large model inference. The NVIDIA GB10 Grace-Blackwell superchip pairs high-performance Blackwell GPU tensor cores with ARM CPU cores inside compact workstation form factors. Low-precision formats like NVFP4 reduce model memory footprints significantly while preserving accuracy close to 8-bit implementations.