By Gregory Sechuga, Program Director, IBM Competitive Insights, gsechuga@us.ibm.com; Nicholas Kunze, Technology Advocate, IBM Competitive Insights, nicholas.kunze@ibm.com; Srirama Krishnakumar, Manager, IBM zSystems Research, IBM Competitive Insights, srirama.k@us.ibm.com

Introduction

Real-time credit card fraud detection and prevention require the ability to check every credit card transaction for fraudulent activity at the time of the transaction using Artificial Intelligence (AI) techniques. A study by Celent[1] found that large banks and payments processors who use AI models for their fraud detection often apply them only on a fraction of transactions due to the negative impact on throughput and latency that having each transaction verified for fraud synchronously. As a result, many fraudulent transactions go unmonitored, undetected, or detected with a delay causing losses by the financial institutions or by the transaction participants. Celent estimates that an average financial institution loses from $18 to $105 million every year due to fraud. In addition to this loss, many legitimate transactions might be declined as they are incorrectly identified as potentially fraudulent transactions. Such experience with delayed approval or rejection of credit cards and other financial transactions might negatively impact customer satisfaction and customer retention.  

The IBM z16 server is the first IBM zSystems platform that is built with the IBM Telum[2] processor. The Telum processor is embedded into the IBM z16 chip and acts as an accelerator for AI. This enables IBM z16 to deliver the speed and scale that is required to infuse AI inferencing into workloads with minimal impact on service delivery.

This paper presents the results of an IBM performance study that compares the performance of a credit card processing application utilizing AI inferencing for fraud detection on an IBM z16 with a built-in AI accelerator with an IBM z15 (previous generation server with no built-in AI accelerator). The study showed that an IBM z16 with an AI accelerator provides four times the throughput and four times faster response time when compared to an IBM z15. This enables real-time fraud detection of all transactions rather than a small fraction as is the case today. The ROI (return on investment) of an upgrade from an IBM z15 to an IBM z16 is estimated to be about 163%% when $18M of cost avoidance (a value typical for a mid-tier financial institution) cost avoidance due to fraud prevention is factored in.

IBM z16 vs IBM z15 performance study for a fraud detection solution

The study compared the performance of a representative credit card processing application on IBM z16 and IBM z15. The application ran on CICS (customer information control system) and was comprised of a combination of COBOL and Java. The COBOL part of the application ran on CICS and communicated with the DB2 database that was running on z/OS. The Java component of the application ran within WebSphere Liberty Server hosted on CICS as well. The Java component provided REST endpoints for accepting credit card transactions and checking for fraud by communicating with the inference service.

Fraud detection was performed by using a scoring service that runs on IBM z/OS container extensions (zCX). The service consisted of an AI inferencing[3] model that ran on Online Scoring Community Edition (OSCE). On both IBM z15 and z16 the scoring service took advantage of IBM zIIPs[4] to reduce the load on the Central Processor (CP) cores and, consequently, lower the client monthly licensing charges (MLC). The client needs to have an adequate number of zIIPs to be able to run the zCX workload. On IBM z16 in addition to leveraging zIIPs, the same scoring service took advantage of the built-in AI accelerator by recompiling the scoring service code and linking it with the IBM zSystems deep neural network (zDNN). This enabled the scoring service to not only perform much better but also reduce the number of IBM zIIPs required. The Java component of the credit card processing application communicated with the scoring service using REST API calls as shown in the diagram below.

The transactions required to drive the applications for measurement of throughput and latency were generated using JMeter[5]. JMeter used 100,000 predefined transactions that were already one-hot encoded[6] to be sent to the credit card processing application. The credit card processing application retrieved 6 additional credit card transactions for the same card from DB2 and then batched the transactions (32 transactions in a batch) before sending them to OSCE that performed inferencing for fraud detection.

Below is a chart that depicts the functional flow of credit card transaction processing.

Infrastructure configuration and comparison methodology

The performance comparison was done by applying hardware and software configurations to IBM z15 and IBM z16 to make them as close as possible by capacity. This was achieved by cloning the IBM z15 LPARs and deploying them onto IBM z16. A total of four CPs and eight zIIPs, a single zCX instance, and two CICS regions with the credit card processing application were used on both IBM z15 and IBM z16.

There was an instance of a JMeter performance test driver for each CICS region. The number of JMeter threads was gradually ramped up to simulate increasing load conditions until reaching saturation of CPs and zIIPs, then the test was driving the same load as the stable system state for a time period.

The following graphs illustrate the performance differences between IBM z15 and IBM z16 with the built-in AI accelerator.

The above graph shows that IBM z16 with the built-in AI accelerator can sustain a throughput that is significantly higher than what can be achieved on an IBM z15.

The above graph demonstrates that an IBM z16 with a built-in AI accelerator not only performs better but also comes with the added benefit of reduced zIIPs utilization. This lowers acquisition costs since fewer zIIPs need to be purchased while buying a new IBM z16 server.

The graph below shows that IBM z16 utilizes up to 85% fewer zIIPs per OLTP transaction with in-transaction fraud detection than IBM z15. This helps the client in deciding on the number of zIIPs they need to purchase while acquiring a new IBM z16.

Return on investment

As mentioned in the Introduction, recent research by Celent showed that a mid-tier financial institution running fraud detection can save at least 18 million US Dollars (USD) per year. Our performance study showed that the IBM z16 with a built-in AI accelerator used in conjunction with zIIPs can process all the transactions in real-time to check for fraud, so the business value of preventing fraud would be at least $18 M per year according to Celent.

ROI assumptions

To calculate an ROI, we assumed a mid-tier financial institution with 10,000 installed MIPS and a typical software stack running on IBM z/OS. We also assumed that the financial institution acquires a new IBM z16. The following tables provide details on the assumptions made to calculate the ROI in this study.

The graph below shows a 3-year TCO comparison of running an already acquired IBM z15 versus a newly acquired IBM z16.

The following tables provide the details on the calculations used to arrive at the return on investment on an upgrade to IBM z16.

The resulting ROI over three years is 163% with an Internal Rate of Return[7] (IRR) of 710%. The graph below further illustrates the business value when viewed in terms of cash flows. As can be seen, the fraud-related cost savings alone can be used to finance the purchase of a new IBM z16 with a breakeven date after the first year.

Summary
This performance study demonstrates that upgrading to an IBM z16 with an AI accelerator provides significant technical advantages that result in added business value to financial firms due to the implementation of real-time credit card fraud detection and prevention. Upgrading to an IBM z16 with Telum processor offers a compelling value proposition to financial firms.

Please, note that while the study conservatively estimates the ROI of fraud detection to be 163% over three years for a mid-tier financial firm, the actual results might be significantly higher: we did not quantify and include the value of cost avoidance related to higher customer satisfaction and retention, and the intangible benefit of the firm’s improved reputation.

References and comments
[1] Katkov, Neil. (2022). Operationalizing Fraud Prevention on IBM z16: https://www.celent.com/insights/667453521
[2] IBM Telum Processor: the next-gen microprocessor for IBM zSystems and IBM LinuxONE:https://www.ibm.com/blogs/systems/ibm-telum-processor-the-next-gen-microprocessor-for-ibm-z-and-ibm-linuxone/
[3] Using Artificial Intelligence (AI) typically consists of two phases – training and inferencing. Machine learning and deep learning refer to training neural networks. AI inference ability is the neural network that enables fraud detection. This study uses only the inferencing capability as the model is already trained for credit card fraud detection.
[4] IBM z Systems Integrated Information Processor (zIIP):  https://www.ibm.com/products/z-integrated-information-processor
[5] The Apache JMeter™ application is open-source software designed to load test functional behavior and measure performance: https://jmeter.apache.org
[6] One-hot encoding is a method of converting data to prepare it for an algorithm and get a better prediction: https://en.wikipedia.org/wiki/One-hot
[7] The Internal Rate of Return (IRR) is a discount rate (https://www.investopedia.com/terms/d/discountrate.asp) that makes the net present value (NPV) of all cash flows equal to zero in a discounted cash flow analysis. The higher the IRR, the more desirable is the investment. In this case, the cash flows are the business value outcomes by saving on fraud prevention.

Acknowledgments
Special thanks to William Lamastro – Principal Infrastructure Architect, Poughkeepsie Competitive Center – for helping us with the design, configuration, test execution, and review of this paper.