General purpose CPUs are dead. Long live the general purpose CPU.

I think that this is what’s happening in the world today – we are witnessing a shift in computing paradigms. The most obvious of these shifts is the exponential growth of artificial intelligence (AI) computations. As more and more computing applications shift towards AI, computing power has to shift as well.

Most already know that the venerable general purpose CPU is not particularly suited to AI computations. So, many shifted to the use of graphics processors (GPUs) to perform these computations instead. However, for Google, even using these highly powerful GPUs were insufficient for their needs as they have determined that all their data centres around the world would not be able to hold enough GPUs needed to power their AI applications.

Instead of having to double it’s data centre foot-print, Google decided to build its own specialised tensor processor (TPU) that is suited to the types of computations used in AI applications. They managed to prove the power of this investment when Deepmind’s Alpha Go, powered by Google’s TPUs, managed to defeat the world’s best players in Go. While the TPU is producing great leaps and bounds in the AI field, Google has no intention in selling the processor to the masses.

Dissatisfied with merely producing general purpose CPUs, Intel purchased Nervana Systems in 2016 and has announced the release of its own processors to handle AI workloads by the end of 2017. This is a great move because it would mean the general availability of AI processors to the public. I can imagine that the growth of AI software applications will accelerate beginning 2018.

Not to be outdone by both these giants, NVIDIA has also announced the availability of their AI processors e.g. Pegasus – that pack massive computational power into a small package. Smartly, they have decided to focus on autonomous driving applications as they have already had a foot in the automobile market previously. One can imagine future automobiles being powered by NVIDIA processors.

However, while AMD has had a quarter of bumper profits, unless they have an AI processor in their works, they are likely to get into trouble once again. It would not be too difficult for AMD to cobble an AI processor together. Their purchase of ATI in the past seems pretty prescient. While their APUs are a first step, I hope that they will take it all the way with an AI processor for the power user.

Regardless of who takes the lead, one thing is sure – general purpose computing now includes AI workloads. Therefore, these AI processors will soon become the general purpose processors in the near future.

At AESTE, this excites us because processor architecture and design, is becoming sexy again.

Categories: Enlightenment


Nur Hussein · 2017-11-16 at 17:49

So what do AI workloads look like? Lots of vector calculations?

    Shawn · 2017-11-16 at 20:08

    NN seem to do lots of weighted sums. So that would be, floating-point multiply-accumulate operations. Vectors would probably help speed things up too. That’s why GPUs are a better fit than the CPU.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.