Lately, I have been writing extensively about the pandemic and changes that are coming in response to it. But my true interest lies in future technology. I’m am fascinated with advances in artificial intelligence, blockchain, and the internet of things. This blog takes me back to those important topics.

I want to discuss how microprocessors work and how advances in AI will require them to work differently. In a way, our advances in chip technology have allowed machine learning and AI to exist, and now that same technology will have to change in order to allow these disciplines to flourish.

What is a microprocessor?

A microprocessor is essentially the primary computation engine of your computer. It’s also known as the central processing unit (CPU). They were originally designed as a collection of chips with transistors wired individually. This will matter in a second, just bear with me.

Nowadays the components are all integrated on a single chip. They can have multiple cores and millions of transistors. Generally, this means higher computing speeds. These chips have been getting smaller and smaller over the years, as well as the components on them.

If you’re interested in how this actually works check out this article. I’ve gone about as far as my brain will allow. I’m a lawyer, not a computer engineer.

Moore’s Law

Why mention transistors? Because of Moore’s Law. Moore’s Law states that the number of transistors on a microchip will double about every two years. Over time, this was translated to mean that computing power would double about every two years.

In fact, Moore’s Law has come true and those gains were realized over the last 50 years. In some ways, this was a self-fulfilling prophecy because the semiconductor industry pushed for these gains. Regardless of its cause, technology has benefited greatly from this trend.

But some are concerned that the trend will end soon because of shrinking microchips. We simply cannot continue wedging more and more transistors onto smaller and smaller chips. It becomes physically impossible. Again, I’m not going to go into crazy detail about this, but there is an interesting article from MIT Technology Review here. If Moore’s Law comes to an end and we don’t see the gains in computer speed, will advanced computing like AI and machine learning continue to move forward?

Artificial intelligence (and other technological advances) continue to require more and more computing power, which has generally been doable thanks to advances in the size of the transistors and the chips they reside on. With those incremental improvements potentially coming to an end, what will take their place? Can we develop new technology that allows computing speed to continue on its upward trend?

The Transition to New Chip Technology

It’s not just raw computing power that needs to change. It’s also the way that computers ‘think’. Machine learning and AI researchers are moving to systems based on neural networks that rely heavily on parallel computations. This is more like how our brains work and may lead to real developments in the field. Traditional CPUs process computations sequentially, which makes them inefficient to train for neural networks. (see an article from Forbes, here)

So what does this all mean?

It means that we need a new microchip. We need a chip that is not restricted in speed by the number of transistors we can jam on it, and we need a chip that can allow for parallel processing to meet the demand of machine learning experts.

Along comes Cerebras, a company that has built the largest chip ever. It’s large but offers ridiculous computing power. Because everything is located on one chip, they are able to move data more efficiently. That, in turn, allows for incredible parallelization.

Another company, Groq, has a unique view of chip performance. Their goal is to move data in very small batches (specifically a batch size of one), but to do it at ludicrous speeds. In fact, they recorded speeds of one quadrillion operations per second. To put that into context, the computer I’m working on currently can do roughly three billion operations per second. The Groq chip is about 333,333 times faster than my computer or 23 million times faster than the computer that took us to the moon. (I am not an expert on these figures, they could be wrong. But the chip is ridiculously fast)

There are a number of other companies in this space. The Forbes article lists a handful more, but the point is that there is a concerted effort to reimagine the microchip so that computers can keep up with the demands of AI and machine learning.

In a lot of ways, my reference to the moon landing was appropriate. The space race led to many technological breakthroughs beyond merely putting a man on the moon. It’s interesting to see how our push for more powerful AI capabilities might lead to similar breakthroughs.

There isn’t a clear frontrunner, yet. At some point, I expect that we will settle on a unified way to generate the speed and parallelization we need. Until then, these companies will continue to test their capabilities against one another. It should be an interesting time.