NVIDIA Looks to Lower AI Barrier to Entry

    NVIDIA this week announced that a development hub that it hosts, which promises to make it both simpler and less risky to train AI models, is now generally available in North America.

    The NVIDIA Base Command Platform is a hosted environment deployed on NVIDIA DGX SuperPOD supercomputers that have been installed in data centers managed by Equinix. Each SuperPOD is made up of 20 systems. A data science team can rent a minimum of three pods monthly so long as they make an initial three-month commitment. Pricing for a monthly subscription pricing to NVIDIA Base Command Platform starts at $90,000.

    The goal is to reduce the cost of training an AI model, said Stephan Fabel, senior director of product management for NVIDIA Base Command Platform. “We’re lowering the barrier to entry,” he said.

    That approach also serves to reduce the level of risk because organizations are not required to acquire systems based on graphical processor units (GPUs) to prototype an AI model that might never find its way into a production environment.

    In some instances, multiple organizations might decide to employ NVIDIA’s Base Command Platform to collaboratively train an AI model in a multi-tenant environment without requiring one of those organizations to acquire a supercomputer, noted Fabel.

    Once the AI model is trained data science teams can then employ NVIDIA Fleet Command tools to deploy an AI model in a production environment.

    The first provider of a machine learning operations (MLOps) platform to partner with NVIDIA on this initiative is Weights and Biases. NVIDIA, however, also expects to partner with other providers of MLOps platforms as well, said Fabel.

    NVIDIA also expects that multiple cloud service providers will eventually partner with the company to provide similar services. Right now, however, Equinix is the only provider of the hosted service. Data and storage management platforms are provided by NetApp under the terms of an exclusive partnership for NVIDIA DGX SuperPOD supercomputers deployed on Equinix data centers.

    Also read: NVIDIA Extends Scope of AI Software Ambitions

    Supercomputers vs. the Cloud

    One of the first organizations to take advantage of the NVIDIA Base Command Platform is Adobe, which is employing the platform to enable research and development teams to experiment with various AI capabilities that might be added to future releases of its software.

    It’s not clear to what degree data science teams will prefer to rent GPUs on supercomputers provided by NVIDIA versus employing cloud services based on GPUs. Cloud service providers such as Amazon Web Services (AWS), for example, have made it clear they intend to offer a mix of services for building AI applications that will include both GPUs from NVIDIA as well as GPUs that it is building. Large enterprises that build lots of AI models may still opt to acquire supercomputer platforms that they deploy in an on-premises IT environment simply because they want to keep as much control over their intellectual property as possible. Each NVIDIA DGX SuperPOD is designed to scale out in a way that could span as many as 140 nodes, noted Fabel.

    The one thing that is certain is AI will one day be infused to varying degrees within every application. The immediate challenge in pursuit of that goal is reducing the cost of experimentation. The more costly it is to train an AI model the fewer the number of research projects there will be launched. Organizations will be forced to make big bets on a few AI models rather than allowing a proverbial thousand AI flowers to bloom. In fact, very few innovations occur based on the original idea set forth. The bulk of innovations are a direct result of unexpected experimentations that generally only occur when the cost of prototyping a new idea or concept is relatively low.      

    Read next: Best Machine Learning Software in 2021

    Mike Vizard
    Mike Vizard
    Michael Vizard is a seasoned IT journalist, with nearly 30 years of experience writing and editing about enterprise IT issues. He is a contributor to publications including Programmableweb, IT Business Edge, CIOinsight and UBM Tech. He formerly was editorial director for Ziff-Davis Enterprise, where he launched the company’s custom content division, and has also served as editor in chief for CRN and InfoWorld. He also has held editorial positions at PC Week, Computerworld and Digital Review.

    Latest Articles