TechCapsules

AI-summarised brief · reviewed before publication

International Business Machines (IBM) has announced a new line of data centre chips and servers that promise to be more power-efficient than rival offerings and simplify the process of deploying artificial intelligence in business operations. The company has introduced its new Power11 chips, marking its first major update to its "Power" line of chips since 2020. These chips have traditionally competed against offerings from Intel and Advanced Micro Devices in data centres, particularly in specialised sectors such as financial services, manufacturing, and healthcare. IBM's Power systems are an integrated package of chips and software, similar to Nvidia's AI servers. According to Tom McPherson, general manager of Power systems at IBM, the company has used this tight coupling to focus on reliability and security. The Power11 systems, available from July 25, boast impressive reliability and security features. They do not require any planned downtime for software updates, and their unplanned downtime each year averages just over 30 seconds. Additionally, they are designed to detect and respond within a minute to a ransomware attack, where hackers encrypt data and then try to extract a ransom in exchange for the keys. In the fourth quarter of this year, IBM plans to integrate Power11 with Spyre, its AI chip introduced last year. This integration will enable seamless AI deployment for inference, the process of putting an AI system to work in speeding up a business task. McPherson emphasized that IBM does not aim to compete with Nvidia in helping create and train AI systems. Instead, the company is focused on simplifying AI deployment for inference acceleration and helping businesses achieve process improvements. "We can integrate AI capabilities seamlessly into this for inference acceleration and help their business process improvements," McPherson said in an interview, referring to work with early customers. "It's not going to have all the horsepower for training or anything, but it's going to have really good inferencing capabilities that are simple to integrate."