IBM Unveils Chip for Accelerating AI Inference Engines

Share
  • August 25, 2021

IBM revealed a processor this week at the Hot Chips 33 conference specifically designed to include acceleration capabilities for running inference engines created using artificial intelligence (AI) models.

The IBM Telum Processor is the company’s first chip capable of providing that acceleration capability while transactions are processing, says Christian Jacobi, an IBM distinguished engineer and chief architect for IBM Z processors used in mainframes.

Scheduled to be incorporated within systems in the first half of next year, IBM expects the AI inference engine to make it simpler to, among other tasks, identify fraudulent transactions in real-time as transactions are being processed, Jacobi says.

Today, most fraud is not detected until after a transaction is processed. IBM claims billions of dollars in fraudulent transactions that occur each year will be prevented from ever occurring in the first place, as AI models for detecting those transactions are first trained and then implemented via an inference engine running on the IBM Telum Processor.

The Telum Processor’s Design

The 7 nm IBM Telum Processor is based on a deep super-scalar, out-of-order instruction pipeline across eight cores running at more than 5 GHz clock frequency. The cache and chip-interconnection infrastructure provides 32 MB cache per core across 32 Telum chips. The module contains 22 billion transistors and 19 miles of wire on 17 metal layers. The IBM Telum Processor is the first IBM chip with technology created by the IBM Research AI Hardware Center. It also makes use of extreme ultraviolet (EUV) lithography technologies developed by Samsung.

That architecture is achievable because IBM has changed the way it packages chips on a processor to increase density. 

“We are going to a dual chip model,” Jacobi says.

IBM has yet to disclose which specific platforms the IBM Telum Processor will be added to next year. However, IBM has been making a case for mainframes that it says are still relied on to process almost 70% of the world’s production IT workloads. The bulk of those workloads are made up of online transaction processing (OLTP) applications. 

It’s not clear to what degree the IBM Telum Processor might keep organizations from migrating workloads off mainframes. However, the more unique capabilities IBM can provide on its platforms, the less likely it becomes that saving money on IT costs will be a compelling enough idea to replace those platforms that will soon be capable of detecting fraudulent transactions in a way that adds additional profits directly to the balance sheet.  

Use of Telum in Fraud Detection 

It may be a while before data scientists can construct the AI models needed to detect fraud in real-time. But IBM is making it clear that platforms capable of applying AI in real-time will be available next year for everything from fraud detection to loan processing, clearing and settlement of trades, money laundering, and risk analysis. 

Just how much of a dent the IBM Telum Processor might put in those illicit activities remains to be seen, but it’s conceivable that the billions of dollars that are lost today because of illicit transactions could be prevented. Organizations might also save additional costs by eliminating the needs for analytics applications that are employed today to detect illicit activity after the fact.

The post IBM Unveils Chip for Accelerating AI Inference Engines appeared first on IT Business Edge.

Source : IBM Unveils Chip for Accelerating AI Inference Engines