IBM is working on a processor with architecture designed for deep learning processes and AI tasks for detecting fraud in the financial market. This Telum processor will be the core chip for the new generation of IBM Z and LinuxONE systems.
IBM revealed details about the new chip at its annual Hot Chips conference. The chip has been specially developed for deep learning processes in the financial market to be used, for example, to detect fraud during bank transactions. Fraud is still largely detected afterwards. The Telum processor is IBM’s first processor to include an integrated accelerator for real-time processing of AI tasks. In this way, for example, credit card fraud can be detected before a transaction has been processed. The accelerator has six teraflops of computing power.
The chip contains eight cores that run at a clock speed of more than 5GHz. A new cache and chip interconnection infrastructure connects each core to a 32MB level 2 cache and allows the AI accelerator to leverage all the caches to form a 256MB virtual level 3 cache. That new infrastructure can scale up to 32 caches that make up a 2GB cache. The new Telum processor has access to 1.5 times more cache per core than IBM’s previous-generation z15 processor, the brand writes in a blog post. That leads to a significant performance improvement per thread, IBM believes. That in turn is necessary for processing AI tasks in real time.
IBM’s Research AI Hardware Center developed the chip, which was made by Samsung using a 7nm process with euv. IBM worked on the hardware for three years, it writes in a press release. The first system with a Telum processor should hit the market in the first half of 2022. This Telum processor will be the central chip for IBM’s new generation of IBM Z mainframe computers and LinuxONE servers.