IBM has introduced its next-generation Telum II Processor and Spyre AI Accelerators, set to enhance the performance of the latest IBM Z mainframe systems, which are specifically tailored for AI workloads. These upgrades are designed to support a variety of AI applications, including traditional workloads and large language models (LLMs), through a new ensemble method of AI.
### Telum II Processor:
The Telum II processor boasts eight high-performance cores, each running at 5.5 GHz. With 36MB of L2 cache per core and a 40% increase in on-chip cache capacity, the total cache now stands at 360MB. Additionally, the processor drawer offers a virtual level-4 cache of 2.88GB, marking a 40% improvement over the previous generation. This processor also features an integrated AI accelerator that ensures low-latency and high-throughput in-transaction AI inferencing, such as enhancing fraud detection during financial transactions. The enhancements culminate in a fourfold increase in compute capacity per chip over its predecessor.
A new I/O Acceleration Unit DPU is integrated into the Telum II chip, offering a 50% increase in I/O density. This integration aims to boost data handling abilities, thereby enhancing the efficiency and scalability of IBM Z systems, making them adept at managing large-scale AI workloads and data-intensive applications.
### Spyre Accelerator:
The Spyre Accelerator is an enterprise-grade solution designed for complex AI models and generative AI use cases. It features up to 1TB of memory spread across eight cards within a regular I/O drawer. Each card is built to handle AI model workloads on the mainframe while consuming no more than 75W of power per card. Each Spyre AI accelerator encompasses 32 compute cores, supporting various data types, including int4, int8, fp8, and fp16, optimized for both low-latency and high-throughput AI applications.
IBM anticipates that its Z mainframe AI systems incorporating Telum II processors will be available to clients by 2025, with the Spyre AI accelerator currently in tech preview and also expected to be available by the same year.
