IBM at NVIDIA GTC 2025

AI looks like a goldmine waiting to be tapped, but many organizations discover that success hinges on their ability to wrangle their data. In this session, learn how to unlock, accelerate and protect your data with the NVIDIA accelerated computing platform to minimize costs, address ethical dilemmas, and overcome technical hurdles that make sustainable monetization of AI tricky. We discuss how to solve for common AI delivery model challenges and accelerate time-to-value with customer-facing AI workloads that leverage NVIDIA technologies.

To unlock AI’s full potential and speed the pace of innovation for specific business needs, organizations must evolve their IT landscape. As AI becomes woven into the fabric of business applications—from core enterprise to consumer-facing services to the edge—there's an increasing focus on reducing cost and maintaining flexibility of deployment. Through demos and conversations with experts, explore how adopting a hybrid platform approach, paired with the power of NVIDIA technologies, removes barriers to AI by delivering better reliability and performance, operational agility and managed costs, no matter your operating environment.

We analyzed the performance of Triton kernels on NVIDIA A100 and H100 GPUs using Nsight Systems and Nsight Compute, which helped gain insights to performance bottlenecks. Modifications of the Triton kernels, using techniques such as SplitK parallelization and others, result in competitive performance results in the same ballpark as cuBLAS.

The common implementation of AI pipelines today is to bring data to AI. This works well when your dataset is relatively small and co-located. When we look at the next step of AI journey, we know one thing for sure: there will be a lot more data in a lot more locations. The effective way to address this challenge is to push AI processing closer to where the data is. This concept is “AI Content-Aware Storage (AI CASt).” The vision of content-aware storage is to enable intelligent storage to process data for AI applications. In this session, we demonstrate the architecture of AI CAST by leveraging NVIDIA Blueprints and NIMs to accelerate the retrieval-augmented generation (RAG) pipeline by incorporating storage and storage metadata in the Continuous Data Ingest and vector DB management.

Producing model card documentation is a state-of-the-art best practice for advancing AI transparently, and educating enterprise developers and nontechnical influencers like customers, investors, users and policymakers about AI in a clear and uniform manner. Learn how to streamline the creation and management of model cards through the model card generator. Understand how the model card automation engine captures essential metadata during training and from model source code, ensuring comprehensive, consistent and up-to-date documentation for AI models.