Transaction processing is the lifeblood of the modern enterprise. And with the next generation IBM Z processor, these applications will be able to run accelerated Artificial Intelligence (AI) processing directly on the platform, providing real-time analytic insights and enhanced security. There are significant potential benefits when applying AI in situ where the data resides and the transactions are taking place, making it possible to inference every transaction to enrich it with trusted, actionable insights.
The Telum Processor
Before we dive into the on-board AI accelerator, let’s take a quick overview of the chip. The design team completely rearchitected the cache hierarchy to improve performance and the amount of cache available to each of the 8 redesigned Telum cores. These cores are interconnected via an on-die fabric that extends low-latency access across 32 chips. Add in the encrypted memory, and a 8-DIMM redundant array of memory and you have perhaps the most secure highly-available platform for enterprise mission-critical applications.
The Telum Integrated AI Accelerator
Today, Mainframe users have two choices for running AI: on the Z CISC cores, or on a separate server with AI acceleration. The first can work well for machine learning codes, by Deep Neural Networks (DNN) require more performance. Offloading can work well, but incurs significant latency and of course exposes another security surface to intrusion and attacks. What people would prefer, and will soon be able to get, is a fast DNN accelerator on the Z processor itself.
The accelerator operates on tensors (the fundamental data type in a DNN) in user space, with matrix multiplication, convolution, pooling, and activation functions all built in firmware. Using firmware will allow IBM to continue to adapt this accelerator for new models and computational functions. The latency is very low and consistent, providing a real-time AI inference capability to every transaction processed on IBM Z.
The upcoming Telum processor for IBM Z and LinuxONE will provide an important step function to enterprise workloads seeking to take advantage of the benefits offered by the application of deep learning neural networks. Real-time scoring of credit card transactions is a prime example where the addition of an AI accelerator on the Z processor will transform business practices, perhaps enabling virtually all transactions to be analyzed for credit card fraud. In addition to the AI accelerator, the Telum design brings features such as new cache architecture that should improve throughput, perhaps significantly, all while maintaining or improving performance, scalability, security, and availability. We note that this accelerator derives from the work done at IBM Research on AI hardware designs. We believe Telum will be perhaps the most impactful IBM Z processor in recent memory.
For more information, please see our Research Report.