ARM Vs. Intel, Phase Two
From 60,000 feet, ARM’s TechCon and the Intel Developer Forum look remarkably similar. The key message from both is a focus on improving performance in processors while significantly lowering power.
Intel wants a piece of the mobile market so badly it can taste it. And ARM, which is the primary processing engine inside of the iPhone and one of a couple in Android-based devices—MIPS has a stake there, as well—now has set its sights on the data center. The margins in big iron are much, much higher than in a mobile device, and there is almost no limit to how many cores can be sold into these markets. Unlike the mobile market, most enterprise applications can take full advantage of symmetric multiprocessing. More cores equal more performance. More efficient cores mean lower energy bills.
Each company has made incursions into the other’s strongholds. Intel’s chips are now in some tablets and at least a couple smart phones. And ARM’s cores will be in AMD’s processors beginning in 2014.
But there also are some significant differences between the two companies that are worth watching. Intel remains remains rooted in the integrated device manufacturer model, for better and worse. It owns its own fabs and can develop its own processes for optimizing its processors, but it also pays for its own manufacturing equipment, and even helps foot the bill on new lithography and manufacturing processes developed by other companies.
ARM, meanwhile, has emerged as one of the central players in a virtual IDM model, where ecosystem partners must work closely together. The upside is that they can benefit from multi-company research efforts, and they can share the skyrocketing costs. The downside is that this arrangement frequently is slower. In an IDM model, fine-tuning a process is often as simple as walking down the hall to the nearest conference room. In a virtual IDM ecosystem, fine-tuning a process involves the efforts of multiple companies, often in different time zones, using different spoken languages, and with teams of lawyers to contend with.
Still, the price of equipping a new fab at 10nm will be upwards of $10 billion. At 5nm, no one even dares to suggest how expensive this might become. Moving from 300mm wafers to 450mm wafers will help, providing yield is sufficient. But whether it’s enough to offset the rising costs of equipping fabs, including multiple types of lithography, incredibly complex chemistry and overcoming quantum effects at 10nm and beyond is something of an educated guess—coupled with a giant leap of faith and a good dose of luck. This is a huge price for even a group of companies to bear.
A second key difference has to do with focus. Intel, for all its talk about reducing power, is working with an x86 architecture and a legacy of energy-inefficient software applications. It’s not that Intel hasn’t done its part. It’s not even that it hasn’t pushed the envelope in advanced features such as voltage and frequency scaling, dark silicon, burst computing, and near-threshold computing. The company continues to invest in new materials and approaches, and it beat the commercial foundries to finFET production by at least a couple of years.
But Intel is a hardware company. It relies on other companies to develop software, and that software requires backward compatibility, which comes at a steep price in terms of efficiency. It’s not that Intel hasn’t tried to overcome that legacy. It has created its own language for developing multiprocessing applications to take advantage of multicore chips, but success there has been very limited. Programmers still think in serial rather than parallel. What’s left is Windows and Linux, and both are big and power hungry.
ARM, in contrast, is first and foremost a commercial IP vendor. It has enough mindshare that Microsoft saw fit to develop Windows RC for ARM’s processor cores, which may prove important in the enterprise market where energy costs are very large and visible. But ARM’s technology also plays in much more power-constrained markets, which is why its processor cores are in the iPhone and Android devices. And as ARM EVP Simon Segars made very clear at his company’s conference this week, the problems that already have been addressed in the mobile world now will be need to be addressed in the server world.
In the data center, virtualization was a Band-Aid. It was effective at improving server utilization, but it was simply a first step. The next step is to actually reduce the cost of computing itself. That means reducing energy costs while maintaining or improving performance. Whether ARM can truly make inroads in that market is unknown. And whether Intel can improve the software that runs on its platforms is equally unknown. Both companies are extremely aware of the challenges and the risks. They just have a different way of getting to the same goal.
Still, the stakes also reach well beyond just the two key players. In Intel’s case, winning in the mobile world will mean more Intel chips. In ARM’s case, winning in the data center will most likely mean ARM cores in other companies’ chips. So while the message at 60,000 feet looks remarkably similar, much closer to the ground the approaches couldn’t be more different.