IBM Research has officially announced the first SoC designed to train and run deep learning models faster than general purpose CPUs: the IBM Artificial Intelligence Unit (AIU).IBM AIU uses PCIe slot like other similar accelerators. The AIU accelerator developed by IBM will be used to solve the basic problems that the company has also stated. These problems stem from the use of traditional CPUs and GPUs in deep learning and training neural networks, hence the inefficiency here. IBM has designed the AIU accelerator entirely for AI training. Thus, the training of artificial intelligence systems will be carried out more quickly and at the same time efficiency will be provided.
An accelerator designed for artificial intelligence
IBM AIU can essentially be described as an application-specific integrated circuit (ASIC). Designed for deep learning, the accelerator can be programmed for any task. With 32 processing cores and 23 billion transistors, the artificial intelligence processor is based on the design used by the company’s previous Telum processors. The Telum processor is manufactured in the 7nm process, while the AIU comes off the bands in TSMC’s efficient 5nm process.IBM’s goal is to increase the efficiency of artificial intelligence hardware 2.5 times every year. The company had established the AI Hardware Center unit to work only in this field. The main goal, stretching out to 2029, is to train artificial intelligence models a thousand times faster than in 2019.
IBM explains why training AI models is so important in its blog post: “Using AI to classify cats and dogs in photos is a fun academic exercise. However, these will not solve the important problems we face today. We need enterprise-grade, industrial-scale hardware for AI to meet real-world challenges.” Finally, IBM hopes to launch the AIU accelerator in the near future.
1638
