Specifications

Overview

The IBM Spyre Accelerator is IBM's purpose-built AI accelerator ASIC, delivered on a PCIe 5.0 x16 card for the IBM z17 mainframe. Each card integrates 32 AI accelerator cores and 128 GB of LPDDR5 memory for 300 TOPS of performance, with a z17 system supporting up to 48 cards across its I/O drawers to scale generative and agentic AI workloads.

Spyre works in an ensemble with the Telum II processor's second-generation on-chip AI accelerator — a 5 nm, 5.5 GHz, 8-core design with 43 billion transistors and a mainframe-first on-chip Data Processing Unit (DPU). Together they enable more than 450 billion inference operations per day with one-millisecond response, supporting multi-model AI on enterprise data.

For QS Compute customers running mission-critical, regulated workloads, Spyre + Telum II bring low-latency LLM inference and fine-tuning inside the trusted IBM Z environment — the first mainframe fully engineered for the AI age.