IBM Develops Telum AI Chip in Collaboration with Samsung Electronics
Deep learning inference can be applied to large-scale financial transactions
Support for AI project implementation where the data is.


On the 24th, at the annual 'Hot Chips' meeting, IBM unveiled details of the 'IBM Telum' processor, which is designed to help respond to financial fraud in real time by applying deep learning inference technology to enterprise workloads.

The Telum processor, developed on the 7nm EUV technology node, is IBM's first processor to include on-chip acceleration technology to enable the application of AI inference technology during transaction processing, and it is also a product in which Samsung Electronics participated as a technology development partner.

The Telum chip consists of eight processors with a deep super-scalar out-of-order instruction pipeline architecture. Each processor runs at a clock speed of over 5 GHz and is optimized for heterogeneous workloads.

The cache and chip interconnection infrastructure provides 32MB of cache per core and can be scaled to 32 Telum chips. The Dual-Chip Module (DCM) design includes 22 billion transistors and 19 miles of wire on 17 metal layers.

The Telum-based system will be released in the first half of next year.

◇ AI projects can be built and executed where the data is.

IBM Telum is designed to efficiently run applications where the data is. It helps overcome existing enterprise AI approaches that require significant memory and data movement capabilities to process inference.

With Telum, you can run large-scale inference on sensitive transactions in real time through accelerators located close to mission-critical data and applications, instead of calling separate platform AI solutions that can impact performance.

In addition, AI models can be built and trained outside the platform, and the AI models can be deployed to a Telum-based IBM system for analysis to perform inference.

◇ Suitable for all aspects of banking, finance, stock trading, and insurance business

Today, companies use detection technology to catch fraud when it occurs. Due to the limitations of existing technology, this process is time-consuming and requires large-scale computing operations.

This is especially true when fraud analysis and detection are performed on a platform isolated from transactions or data related to a company's core business. Due to latency, complex fraud detection often does not occur in real time. This means that criminals can purchase goods using stolen credit cards before retailers even realize that fraud has occurred.

Telum can help enterprise customers shift their mindset from a passive approach of detecting fraud to an active one of preventing fraud before a transaction is completed without impacting Service Level Agreements (SLAs). The new chip enables the full performance of the AI processor for AI-specific workloads.