For a couple of years, many of Nvidia’s AI announcements were targeted around “democratizing AI,” making it accessible for a large number of enterprises. More recently, with the emergence of generative AI, just having access to large language models is not enough; enterprise developers need to have the ability to customize generative AI technology to leverage the data and needs of their own organizations.
Nvidia made a big splash this week at SIGGRAPH 2023, a conference that caters to computer graphics developers and researchers, announcing at the Los Angeles event a new toolkit that developers can use to quickly create, test, and, perhaps most important, customize pre-trained generative AI models on a PC or workstation. Once customized to their preference, these models can be scaled to any data center, public cloud or Nvidia’s recently unveiled DGX Cloud.
Erik Pounds, senior director of enterprise computing at Nvidia, told Fierce Electronics that the toolkit, called Nvidia AI Workbench, helps enterprise developers deal with a complex array of challenges that are “not for the faint of heart,” such as choosing and working with the right software and components, many of them open source in nature, including repositories and models from communities like Hugging Face and GitHub.
“What Workbench helps you to do is assemble your project right from your local machine, right from your PC, your laptop, or your workstation, and then be able to run that project anywhere you need to run it,” Pounds said. “So whether that is on your local machine, assuming you have a GPU that has enough memory and horsepower to do what you're trying to do, or in your local data center, your company's data center you have access to or up in the cloud.”
Pounds added that AI Workbench also helps developers who just are at the experimental stage with generative AI. They likely are seeing many new models coming out, and want to see how easy it is to customizable a model and see how it responds before they decide about moving forward with a project. In these cases, they can use Nvidia’s NeMo framework to train the models, particularly as their confidence grows and they want to train a model on larger data sets. Nvidia also announced at SIGGRAPH the latest version of its enterprise AI software platform, Nvidia AI Enterprise 4.0, which includes the latest version of the NeMo framework, among other tools.
AI Workbench addresses the notion that “not every not every organization or company spent the last five years building an AI team,” Pounds said. While companies often have application teams, data science teams, and large groups of engineers to work on AI projects, “sometimes they don't have the full expertise to get a production application into deployment, so anything we can do to make that easier will help,” he said.
Nvidia’s announcement comes at a time when the number of generative AI large language models continues to explode. There are somewhere in the neighborhood of hundreds of thousands of such models that developers can choose from. For example, users of the Hugging Face community have shared more than 250,000 models and 50,00 data sets, according to Nvidia.
With that in mind, Nvidia also announced at SIGGRAPH that it is partnering with Hugging Face to give developers access to DGX Cloud within the Hugging Face platform to train and tune advanced AI models.Pounds explained, “We're integrating Hugging Face into DGX cloud so you can go into Hugging Face, find a model, and continue customizing and training that model all within a few clicks straight from the Hugging Face UI right into Nvidia DGX cloud and also into the AI Workbench workflow.”