Abstract:
On October 21, NVIDIA released the modular AI factory reference design DSX POD at the GTC Berlin conference. This solution involves partners pre-integrating the Vera Rubin system and supporting infrastructure into modular containers and completing verification before leaving the factory. NVIDIA said that under the premise that the site has been prepared, the deployment time can be shortened from six months to as fast as 14 days.
DSX POD moves the integrated installation of compute, networking, cabling, rack-level power and liquid cooling systems to the manufacturing facility. Partners complete inter-rack cable connections, electrical wiring, pipeline installation and functional verification before delivery, and then access the site's power, cooling, network and other supporting facilities after delivery.
According to NVIDIA's introduction, equipment pre-integration and site preparation can be carried out separately to shorten the time required to put new computing power into operation. Completing testing in the factory also helps reduce the impact of on-site construction errors, parts delays and complex logistics arrangements on the deployment schedule.
The fastest 14 days refers to the deployment time on the prepared site, and does not include the entire cycle of the AI factory from site selection to completion. Customers still need to prepare power, cooling, network and other support facilities in advance.
NVIDIA founder and CEO Jensen Huang said that the node where the AI factory begins to generate value is when it is put into operation and generates Tokens. He believes that shortening the deployment cycle can help customers launch model training and production environment inference services earlier, and strive to obtain revenue in advance.
NVIDIA plans to work with computing and infrastructure partners such as Dell Technologies, HPE, Lenovo, Bull, Schneider Electric, and Vertiv to deliver modular AI factories based on DSX POD design globally in early 2027.
Comments