ARM announced Tuesday that it will let customers customize their own chips for smartphones and machine learning jobs with the goal of achieving a boost in sustained peak performance without sacrificing too much battery life. The program will let customers work closely with ARM engineers to adjust the ARM IP for specific workloads. The move signals a willingness by ARM to adjust its traditional IP licensing models to meet the changing reality of the silicon market and is made possible in part by the huge demand of custom chips for new applications.
ARM calls the new program the Cortex-X Custom and its first chip designed for this effort is the Cortex-X1. For chipheads, ARM provides a lot of the details about the architecture and program on its blog. To be clear, while ARM calls this a custom core, it’s really just a way to customize ARM’s existing silicon cores to optimize specific features based on the job the chip needs to do. For a truly custom chip, ARM licensees still need to get an architectural license.
To understand this move, it’s important to understand what ARM has done since its founding 30 years ago.
The company came on the scene with an architecture optimized for lower power usage. In comparison, other architectures such as Intel’s were designed for processing a lot of data quickly with little regard to power optimization. Intel’s focus on power wouldn’t come until data centers demanded Intel deliver their need for performance without requiring a new power plant for every server farm.
Because of that focus on efficiency ARM’s designs were perfect for battery-powered smartphones, and the company ended up in almost every handset sold. In the late aughts as the data center operators demanded more power-efficient chips ARM started to develop a program for higher-performance computing as well. The company also saw opportunities in graphics chips as mobile phones became capable video players and gaming machines, so it designed Mali graphics chips for the embedded device market.
The point here is that ARM adapts to the environment, especially if that environment needs to optimize for low-power without sacrificing too much performance. And the internet of things provides a bonanza for ARM, with sensors, AR headsets, phones and other battery-powered devices (or simply devices that don’t need to suck down a watt when half a watt will do). ARM’s challenge has been that it had to do this with a fairly limited number of designs. When you design chips you need to find a wide-enough market for your product to make the years of development effort worth it, or you need to design a specialty chip with a lucrative enough market that you don’t have to spread those high costs out over a lot of sales.
In the data center, we saw companies such as Google, Microsoft, and Amazon realize that their millions of servers represented a large enough market. Additionally, their needs were so specific, that they could start designing chips for some of their cloud computing jobs for which a large investment made sense. Apple realized the same thing when it built a specialized chip for the iPhone.
Essentially, if you have a product that sells hundreds of millions of units and has highly specific needs, it may make sense to design a chip for those needs. But there’s also a growing market of products with a large number of sales that can’t quite justify the expense of a custom chip, yet would benefit from a little more optimization.
And that’s where ARM’s new Cortex-X chips come in. ARM’s telling the market that it can adapt to this world and that it thinks there are multiple use cases where unit sales will justify the cost of a chip that’s highly optimized for the jobs that the device needs to perform. This is great news for the tech industry.
It’s a measure of ARM’s faith in the future of AR/VR, machine learning at the edge, and devices that we don’t know we need yet. It’s also an admission that the future is going to be so varied that ARM can’t predict where it will go, and so it wants to make those bets with the people building it. I’m stoked to see what chip companies and product companies build when they can work early and often to build silicon for the future.