Nvidia points to the future of AI hardware

Unmatched power for AI

For CIOs with AI aspirations, the Blackwell announcement will signal the ability to experiment with super chips or dedicated servers, Nguyen adds. Blackwell will allow enterprises with major AI needs to deploy so-called superpods, another name for AI supercomputers. Blackwell will also allow enterprises with very deep pockets to set up AI factories, made up of integrated compute resources, storage, networking, workstations, software, and other pieces.

The case for Blackwell is clear, adds Shane Rau, research VP for semiconductors at IDC. As AI models get larger, they’ll require more performance for training and inferencing, the process that a trained AI uses to draw conclusions from new data, he says.

As LLM AIs trained in 2023 are deployed, “CIOs will learn what works and what doesn’t, and so a retrain and redeployment cycle will begin,” Rau says. “Thus, the need for Blackwell should be strong.”

If organizations aren’t training their own LLMs, the AI case for Blackwell is highly dependent on their industry verticals and internal workflows, Rau adds. “The more application-specific the workload they have and fewer resources they can bring to bear, the longer they’ll have to wait for AI solution stack and AI model standardization,” he says.

NIM, Nvidia’s software package to optimize inference for several AI models, should also gain traction in the market, because many companies won’t be able to train AIs for their purposes, Rau says.

“Not everyone has the resources to train and deploy AI models at scale, nor do folks want to buy general models when all they need is a model specific to their identified workloads,” he says. “So, pre-trained models and run-time models made off-the-shelf for IT folks to buy and maybe tune a little bit, will be necessary for AI to scale across enterprises and across the internet.”