NVIDIA has introduced the DGX Spark 64GB, a desktop system designed for local AI agent operation, fine-tuning, and inference. The device features the GB10 Grace Blackwell superchip, which combines a Blackwell GPU with 5th-generation Tensor Cores and a 20-core Grace Arm CPU, delivering up to 1 petaFLOP of FP4 AI compute.
The system is equipped with 64GB of unified LPDDR5x memory, utilizing NVLink-C2C technology for shared access between the CPU and GPU. This architecture eliminates the need to copy weights between system RAM and VRAM. The unit includes ConnectX-7 networking at 200GbE, allowing users to cluster up to four systems to scale memory and compute capacity.
The DGX Spark 64GB runs on standard power outlets without specialized cooling. It ships with the NVIDIA AI software stack, including DGX OS, PyTorch, Jupyter, and Ollama. The system supports local execution of open-weight models, such as the Muse Glimmer model, and enables QLoRA fine-tuning on 70B parameter models.
NVIDIA provides the Sync Cluster Assistant for management, which supports configuration across local networks and Tailscale meshes. According to company data, clustering two 64GB units provides up to 1.7 times the performance of a single 128GB DGX Spark, with combined bandwidth reaching 546 GB/s.
