IBM Telum

IBM Telum is the codename for IBM's first processor with an on-chip accelerator for AI inference while a transaction is in process. Announced in August 2021, and commercially available in the IBM z16 system in April 2022, Telum is designed to accelerate deep learning inference, particularly for real-time fraud detection and other transaction-intensive workloads.

The processor is built using a 7nm process technology and features eight processor cores, each with a deep out-of-order instruction pipeline. A key innovation is the integration of a dedicated AI accelerator directly onto the processor chip, allowing for significantly faster AI inference compared to relying solely on CPUs or external accelerators. This close proximity and integration minimize latency, a critical factor for applications requiring immediate decisions based on AI analysis.

Telum's on-chip AI accelerator supports the acceleration of inference for various deep learning models, with a primary focus on models used in financial services and other industries where rapid transaction processing and fraud detection are paramount. The integration of AI capabilities directly into the processor allows IBM z16 systems to perform AI-powered tasks without the need to move data off-chip, thus improving performance and reducing latency.

The architecture also incorporates a dual-chip module design, combining two Telum processor chips into a single package. This allows for increased core density and improved overall system performance.