NVIDIA announced a 64GB unified memory configuration for its DGX Spark personal AI supercomputer. Available this month through hardware partners, the system utilizes Grace Blackwell compute and the NVIDIA AI software stack to run 100-billion-parameter models and agentic applications locally without cloud dependency.
Local agent development hardware
The new DGX Spark SKU integrates NVIDIA Grace Blackwell compute with 64GB of unified memory and ConnectX-7 networking. It is designed to host AI agents, inference, and fine-tuning tasks on-device. The platform arrives with DGX OS pre-installed, allowing researchers to process data science workloads locally rather than relying on cloud instances for every experiment.
A dedicated clustering feature enables two 64GB units to link via the NVIDIA Sync Cluster Assistant. This software automates network configuration and device validation. In internal testing using a Qwen model, a two-unit cluster achieved up to 1.7x performance compared to a single system, providing a scaling path for larger local workloads.
Availability and system constraints
This configuration will be released through manufacturing partners including Acer, ASUS, Dell, Gigabyte, HP, and MSI. While it retains the core software stack and Superchip architecture of the 128GB version, the 64GB capacity serves as a more accessible entry point for edge development.
The reported performance scaling results are based on NVIDIA's specific tests with a 27B parameter model. Real-world efficiency in multi-node clusters may vary depending on the specific agentic application or fine-tuning task. Access is currently limited to the specified manufacturer partners starting later this month.
Первоисточник
В этом отчете кратко изложен материал указанного ниже источника. Аналитика приводится отдельно; заявления о продуктах и исследованиях атрибутируются их авторам.
Читать оригинал на NVIDIA