NVIDIA Unveils "NemoClaw": Empowering Local AI Personal Assistants on DGX Spark and DGX StationNVIDIA has officially launched NemoClaw, a sophisticated personal assistant software designed to run locally on its high-performance workstations, including the DGX Spark and DGX Station. This initiative aims to provide users with a powerful, private AI assistant without the recurring costs of cloud-based token fees.
Breaking the VRAM Barrier
A significant hurdle for local AI deployment is the limited VRAM (Video RAM) on consumer-grade hardware, which often restricts users to smaller models. To solve this, NVIDIA announced a major software update that allows users to link up to four DGX Spark units seamlessly. This configuration provides a massive combined memory pool of 512GB, more than enough to handle large-scale LLMs (Large Language Models) locally.
The Return of the Supercomputer-in-a-Box
For users requiring even more power, NVIDIA is positioning the DGX Station, powered by the NVIDIA GB300 (Grace Blackwell) chip. Featuring a staggering 748GB of unified memory, a single DGX Station is capable of running a 1-Trillion parameter model (quantized to FP4).
While the DGX Station was first teased earlier this year, NVIDIA has now confirmed it is officially accepting orders, with shipments expected to begin in the coming months. Leaked retail pricing for the DGX Station is approximately $97,000.
The key to the DGX Station's ability to run 1 trillion (1T) models is its use of the NVFP4 data format, NVIDIA's new standard that offers twice the performance of the older FP8 in terms of processing speed and memory management. This makes running large models on a single machine at home or in the office a reality.
NemoClaw isn't just about saving on token costs; it addresses the "data privacy" concern for organizations and individuals worried about sending data to other companies' clouds. Having a sophisticated GPT-4 AI running under a desk without an internet connection is a dream for lawyers and financial institutions.
While the price tag of nearly $100,000 may seem high, for AI startups, the DGX Station is "cheaper" than long-term cloud leasing. A one-time purchase allows development teams to train and fine-tune models 24/7 without hidden costs.
The system, connecting four DGX Spark machines, utilizes a new generation, ultra-high-bandwidth C2C (Chip-to-Chip) Link technology. This allows the system to see 512GB of RAM as a single unit (Unified Memory), eliminating the bottleneck issues experienced with traditional connections.
Meta Bets $27 Billion on Nebius to Power the Next Generation of AI Factories.
Source: NVIDIA
NVIDIA Unveils "NemoClaw": Empowering Local AI Personal Assistants on DGX Spark and DGX StationNVIDIA has officially launched NemoClaw, a sophisticated personal assistant software designed to run locally on its high-performance workstations, including the DGX Spark and DGX Station. This initiative aims to provide users with a powerful, private AI assistant without the recurring costs of cloud-based token fees.
Breaking the VRAM Barrier
A significant hurdle for local AI deployment is the limited VRAM (Video RAM) on consumer-grade hardware, which often restricts users to smaller models. To solve this, NVIDIA announced a major software update that allows users to link up to four DGX Spark units seamlessly. This configuration provides a massive combined memory pool of 512GB, more than enough to handle large-scale LLMs (Large Language Models) locally.
The Return of the Supercomputer-in-a-Box
For users requiring even more power, NVIDIA is positioning the DGX Station, powered by the NVIDIA GB300 (Grace Blackwell) chip. Featuring a staggering 748GB of unified memory, a single DGX Station is capable of running a 1-Trillion parameter model (quantized to FP4).
While the DGX Station was first teased earlier this year, NVIDIA has now confirmed it is officially accepting orders, with shipments expected to begin in the coming months. Leaked retail pricing for the DGX Station is approximately $97,000.
The key to the DGX Station's ability to run 1 trillion (1T) models is its use of the NVFP4 data format, NVIDIA's new standard that offers twice the performance of the older FP8 in terms of processing speed and memory management. This makes running large models on a single machine at home or in the office a reality.
NemoClaw isn't just about saving on token costs; it addresses the "data privacy" concern for organizations and individuals worried about sending data to other companies' clouds. Having a sophisticated GPT-4 AI running under a desk without an internet connection is a dream for lawyers and financial institutions.
While the price tag of nearly $100,000 may seem high, for AI startups, the DGX Station is "cheaper" than long-term cloud leasing. A one-time purchase allows development teams to train and fine-tune models 24/7 without hidden costs.
The system, connecting four DGX Spark machines, utilizes a new generation, ultra-high-bandwidth C2C (Chip-to-Chip) Link technology. This allows the system to see 512GB of RAM as a single unit (Unified Memory), eliminating the bottleneck issues experienced with traditional connections.
Meta Bets $27 Billion on Nebius to Power the Next Generation of AI Factories.
Source: NVIDIA
Comments
Post a Comment