AI Accelerators: The Future of Computer Hardware?
As AI technology advances, processor performance requirements have also advanced; we now require faster processors than ever before to cope with the demands of the programs we’re creating. Plenty of people have forecast an imminent end to Moore’s Law; we could soon approach the absolute limit of miniaturisation. This begets the question: how could we continue to improve processor performance for running AI programs? Well, the answer may lie with AI accelerators.
The concept of AI accelerators, special processors for running AI systems, has been around since the early 1990s. However, it wasn’t until recently that they’ve started to come to fruition. With the rapid advancement of AI technology, and the large quantity of computation required to utilise this technology, companies are desperate to improve their AI systems in whatever ways they can, and maximising efficiency in running AI programs appears to be a great way to do that. It is therefore no surprise that AI-focused processors are all the range at the moment.
IBM Watson has explored a number of different AI accelerators. Their digital accelerators reduce processing speed for AI programs by designing hardware specifically for operations commonly used in deep learning; such as the dot product. IBM have also manufactured Analogue accelerators, which drastically increase the speed of learning in image analysis tasks.
Recently, Google announced that they were making the fourth version of their Tensor Processing Unit (TPU). Each chip contains two cores, which in turn consist of MXUs (matrix units). The presence of the MXUs allows the TPU to specialise in processing linear algebra; enabling it to efficiently compute vast matrix and vector calculations. The TPU runs on a cloud, and uses High Bandwidth Memory to communicate with the device being used. The specifications of the latest iteration are not yet known, however the third iteration contained an astonishingly large 4 TebiBits (2^40 bytes) of memory and 8 cores. These accelerators have been used within Google for years now, powering services like Gmail and Translate.

The Cloud TPU v2
Credit: Google
Apple has also been driving forward advancements in AI accelerators. The Apple A13 Bionic, which appears in the iPhone 11 generation of phones, has a 6 core CPU. Within this CPU, there are machine learning accelerators called AMX blocks, which specialise in matrix calculations and can perform up to 1 trillion operations per second. Such powerful hardware allows Apple devices to function faster than ever before, while also leading to an increase in battery life, rather than the battery life reduction one would expect.

The Apple A13 Bionic features astonishing specifications, including a dedicated ML controller and ML accelerators.
Credit: Apple
Certainly, AI accelerators seem to be the way forward. As our data gets larger and larger; our AI programs increasingly complex, it is clear that dedicated AI-focused processing seems to be the best way to ensure efficiency when running AI programs. Who knows, in 10 years, AI accelerators could be standard hardware for every device, much like Graphics Processing Units have become a staple of household PCs.
Thumbnail Credit: geralt