The large image: After trying (and failing) to place Watson as the following era platform for AI functions, IBM is now specializing in creating {hardware} parts for the newest generative AI fashions. The market is evolving, AI know-how is shifting into manufacturing, and Large Blue is keen to say a share of Nvidia’s dominance sooner reasonably than later.
IBM lately introduced the Telum II Processor and the Spyre Accelerator, two chip designs geared toward aiding clients with trendy AI workloads. The company, naturally, prioritizes promoting its personal {hardware}, which is why each chips are solely appropriate with IBM z16 mainframe computer systems.
Telum II is the newest iteration of the Telum structure, launched in 2021. IBM acknowledged that the brand new chip was developed utilizing Samsung’s 5nm manufacturing course of and options eight high-performance cores operating at 5.5GHz. The corporate additionally revealed a 40 p.c improve in on-chip cache reminiscence, with digital L3 and L4 capacities increasing to 360MB and a couple of.88GB, respectively.
The Telum II chip additionally features a novel information processing unit, designed to speed up I/O operations straight inside the CPU. “These {hardware} enhancements are designed to offer important efficiency enhancements for shoppers over earlier generations,” IBM acknowledged. Every new Telum II processor is predicted to ship a 4x improve in computing energy, reaching 24 trillion operations per second (TOPS).
TOPS alone do not inform the entire story, IBM acknowledged. The Telum structure has been improved and optimized for at this time’s AI ecosystem, with excessive throughput and low-latency inferencing. The brand new chip additionally helps INT8 information varieties, which ought to improve effectivity in functions designed with INT8 know-how, reminiscent of newer AI fashions.
The second piece of AI {hardware} launched by IBM at Scorching Chips 2024 is the Spyre Accelerator, a PCIe card containing 32 AI accelerator cores, which share the same structure to the AI accelerator included within the Telum II processor. IBM means that potential clients use each the Telum II and Spyre to run bigger AI mannequin units in what the corporate calls “ensemble AI” use instances.
The ensemble AI methodology leverages a number of AI fashions to reinforce efficiency and accuracy within the remaining outcomes. IBM defined this know-how utilizing a claims fraud detection instance, the place the preliminary danger evaluation made by conventional neural networks is mixed with giant language fashions. In accordance with IBM, ensemble AI methods are so efficient at optimizing AI workloads that they will adjust to regulatory necessities whereas mitigating monetary crimes.
The Telum II processor and Spyre Accelerator have broad use instances. IBM highlighted that its new chips can assist fraud detection, superior anti-money laundering fashions, and extra. They can be used to develop AI assistants, the corporate added.
