While the word “mainframe” may conjure up images of reels of magnetic tape and green symbols on a terminal screen, these systems remain key to big business: 71% of Fortune 500 companies use mainframes for mission-critical tasks.
The mainframe market was worth $5.3 billion last year, so it’s a juicy piece of the action, and IBM, the main manufacturer of these systems, isn’t going to let it slip away. Today, it announced the z17 mainframes, which replace the z16 generation and offer much greater performance in today’s AI scenarios.
Source of images: IBM
The new platform is based on the Telum II processors with z/Architecture announced last fall. This generation of chips uses the Samsung 5HPP process technology and includes eight improved Telum cores with a frequency of 5.5 GHz with a redesigned and significantly strengthened caching subsystem.
The new trends have fully affected Telum II: each processor received support for INT8/FP16 formats, as well as an AI coprocessor with a performance of 25 TOPs (768 TOPs for a fully configured mainframe). By today’s standards, these numbers do not look too high, but in the z17, Telum II processors will work together with 32-core Spyre AI accelerators.
At the time of the announcement, we are talking about 48 Spyre chips in each z17 mainframe, which at 300 Tops per accelerator gives a total of 14.4 Pops (petaops), but within a year it is planned to double this figure – up to 96 accelerators per mainframe, which, taking into account the capabilities of Telum II itself, will already give 30 Pops. The DPU block integrated into Telum II will help the platform digest the volumes of data typical for modern AI models.
According to the announcement, the z17 can process up to 450 billion inference operations per day, which is 50% more than the previous z16. IBM spent more than 2,000 hours designing the z17, researching feedback from more than a hundred customers. The company considers the new product to be a universal solution with more than 250 use cases for AI tasks, including as a platform for generative AI.
The first z17 deliveries will begin on June 18, but without the Spyre accelerator boards yet — the latter will debut later, in Q4. IBM is also preparing a new operating system for the new platform, z/OS 3.2, which should debut in Q3. It will have native support for hardware AI accelerators, as well as AI controls for the system itself. In addition, z/OS 3.2 will implement support for new types of data access, NoSQL databases, and hybrid cloud computing.