IBM has developed a new artificial intelligence (AI) application specific integrated circuit (ASIC) containing 32 processing cores and 23 billion transistors to run any deep learning task.
Called the AI unit (AIU), the chip is designed for deep learning, whether it's processing spoken language, words or images on a screen. The IBM AIU is designed to be an easy-to-use graphics card and can be plugged into any computer or service with a PCIe slot.
IBM said it did not design the chip from scratch but used a scaled version of its AI accelerator built into its Telum chip. The 32 cores in the IBM AIU resemble the AI core embedded in the Telum chip that powers IBM’s z16 system.
One major difference is that Telum uses transistors that are 7 nm in size while the AIU will feature faster and smaller 5 nm transistors.
Why it is needed
Prior to the birth of AI, central processing units (CPUs) were the power behind most forms of computing. While suited well for general-purpose software applications due to their flexibility and high precision, these same qualities put it at a disadvantage when it comes to training and running deep learning models that require massively parallel AI operations, IBM said.
IBM said that running CPUs or GPUs for deep learning models may get the job done in the end, but it is not the most efficient or best way to do it. An all-purpose chip optimized for the types of matrix and vector multiplication operations for deep learning is what was needed.
The AIU architecture features a simpler layout than a multi-purpose CPU and has been designed to send data directly from one compute engine to the next, which creates energy savings.
The goal
IBM said the goal is to train and run AI models 1,000 times faster than previous years and expects to reach this level by 2029.
The company said that with the time and expense involved in training and running deep-learning models, what can be done with AI has barely scratched the surface.
The next few years will be deploying AI to tackle complexities in the real world with enterprise-quality, industrial-scale hardware, IBM said.
