Recently, IBM announced the release of their latest mainframe designed to tackle fraud with the use of AI. What challenges do typical mainframes face when handling transactions regarding fraud, what features does the Z16 have, and is the media misleading the public with claims of the Z16 integrating the industry’s “first on-chip AI inference engine”?

Why is fraud a big challenge for mainframes handling transactions?

One would be forgiven for believing that the computers we see and use every day are responsible for everything we do, whether paying for shopping, making changes to contacts in a contact book, or even changing a subscription to a service. However, most of the machines that we interact with directly on a daily basis are rarely responsible for any exchange of information that needs to be checked for integrity, stored, and processed by a company.

For example, a shopping cart on a website will present the user with a price for the total value. Now, a user could very easily open the source of the webpage, change the price, and have a site that shows a different price. However, upon going to the checkout, the website would not read the value of the user’s webpage but would instead calculate the value itself internally on a server, and this would then be presented to the user.

If, however, a user was able to confuse a system on the price of goods, then this would be counted as fraud, and it turns out that fraud is a very real problem that affects computers daily. According to the Federal Trade Commission (FTC), Americans lose more than $5.8bn a year due to fraud, whether it is falsifying records, stealing personal information, theft of credit cards, and intentional underpayment. But fraud does not just include money; it can also be falsifying information (such as age), falsifying credentials (used to get around credit check agencies), and misusing resources.

Fraud can be difficult to spot at the easiest times, and mainframes responsible for processing transactions can easily miss such fraud. It should be noted that the term transaction does not just refer to the exchange of money; any movement of data can be considered a transaction. For example, an exchange of two contracts between two parties would be a transaction, the allocation of resources to a user would be regarded as a transaction, and even updating data in a database by a user could be considered a transaction.

Such fraud is hard to determine as identifying suspicious activity requires active monitoring of each transaction, looking up the history of the parties involved, and then deciding if the transaction is something that would be expected. For example, two businesses may frequently exchange goods and services with each other, so any transaction between these two is expected. However, if one of these businesses gives goods and services to the other through a third party out of nowhere, this could be flagged as suspicious. Using traditional computing methods, this type of algorithm is extremely hard to implement efficiently when a mainframe is expected to process millions of transactions every second.

IBM announces Z16 mainframe to tackle fraud

Recently, IBM announced that its current line of mainframes, the Z15, are being replaced with the Z16, which has been designed with fraud prevention at its heart. The new mainframe integrates the IBM Telum Processor that itself incorporates 8 IBM z/Architecture cores, 32MB L2 cache, an out-of-order pipeline, and a maximum core frequency of 5GHz.

However, what makes the Telum processor of particular interest is that it also integrates an AI inference engine that runs simultaneously as the processor and can be used to monitor transactions in real-time. This use of AI will allow for suspicious activity to be picked up while it is being calculated, and this can then be used to flag security experts of potential fraud activity.

Furthermore, IBM has also stated that their Z16 mainframes have been built to be quantum-safe, meaning that future quantum computers would still not be able to easily decrypt encrypted data in the mainframe. This helps future proof the Z16 mainframe against future attacks. However, when considering how quantum computers are still yet to become practical, this may not be a feature that is readily needed in modern computing. Of course, when dealing with financial transactions, having the most secure system is a must.

IBM has also announced that they will be providing updates to the Watson Machine Learning for IBM z/OS platform to allow businesses to better analyse data in real-time and provide insights into operational data.

Is the “first AI chip” claim misleading?

There seem to be multiple media sites that state that the Telum processor used in the Z16 is the industry’s first AI inference engine on a chip. However, as the media often does, this claim is only half-true as anyone from an engineering background would immediately think to themselves of the Snapdragon SoCs, which integrate AI inference engines and Coral computers, which have a Tensor accelerator core.

What the IBM Telium Processor does that is unique is that its AI inference engine, which is on-chip, operates on transactions that are being processed simultaneously. This means that instead of processing transactions and then using the AI engine to observe them, the AI engine does this in parallel with the main core so that fraud can be detected as it happens.

The Z16 is an impressive mainframe, and its ability to detect fraud in real-time will be a valuable asset in the fight against fraud, but to call it the industry’s first AI on a chip is wrong.