Hybrid AI on IBM Z: Options and Technical Insights

Abstract

The rapid growth of AI (Artificial Intelligence) and ML (Machine Learning) over the last few years has enabled organizations to seamlessly integrate real-time decision models into their existing applications for more efficient processing. In response to this growth in AI, IBM has increasingly built upon its mainframe architecture most notably IBM Z by incorporating newer technologies like support for Python, containerized workloads through zCX, and virtualization layers like z/VM and zKVM. The introduction of the Telum processor, with built-in AI acceleration, further positions IBM Z as a strong candidate for running AI workloads right where the data resides. This paper explores several options that can be used to deploy ML model training and inference on IBM Z's ecosystem. It highlights how IBM Z's performance, security, and co-location with enterprise data make it an ideal environment for hybrid AI workloads. This paper takes an example use case of real-time credit card fraud detection and explores how predictive models can be deployed on IBM Z using Python. This paper also explores the full transaction flow to reflect the integration between existing COBOL, HLASM, or Java applications and how the Python-based fraud scoring service would work in practice