NVIDIA Announces DSX Flex to Optimize AI Factory Power Consumption
NVIDIA has unveiled DSX Flex, a grid-orchestration platform designed to optimize AI factory power consumption, enabling more efficient use of energy and better grid compatibility.
On a sweltering August evening in Silicon Valley, NVIDIA demonstrated the capabilities of its new DSX Flex platform, which helps AI factories adjust their power consumption based on grid conditions. During the demonstration, the AI factory received signals from Silicon Valley Power to adjust its power usage without any manual intervention.
The DSX Flex platform, developed in collaboration with NVIDIA partner Emerald AI, allows data centers to reschedule less urgent computing tasks when grid conditions require it, ensuring that higher-priority services continue to operate smoothly.
NVIDIA plans to deploy DSX Flex commercially at its 96-megawatt Vera Rubin AI factory in Manassas, Virginia, following five prior demonstrations across two continents. This deployment will serve as a model for future AI factory designs.
In addition to DSX Flex, NVIDIA is incorporating an 800V DC power architecture into its reference designs to improve power delivery efficiency and support denser accelerated computing racks. This architecture is designed to reduce conversion complexity and support the scaling of AI factories.
NVIDIA DSX, the company's comprehensive suite of tools, includes DSX Sim, DSX OS, DSX Exchange, and DSX Reference Designs. These tools are designed to optimize the entire AI factory, from simulation and operational software to power management technology, ensuring that the factory produces more useful work per megawatt consumed.
NVIDIA claims that its DSX suite can increase cluster token throughput by 24% and performance per watt by 23% compared to traditional setups, demonstrating the potential for significant efficiency gains in AI factory operations.
Source: nvidia-newsroom