When Will It Be Done?

Design teams have done remarkably well in getting chips out the door on time, despite growing complexity at each new node and an increase in the number of features and IP blocks that need to be integrated into designs.

There has been plenty of grumbling, along with dire warnings about the future of Moore’s Law and the impact of industry consolidation. The reality, though, is that the volume of chips shipped each year is going up. In December, the Semiconductor Industry Association predicted 3.3% growth in 2017, or roughly $346 billion in sales, with another 2.3% growth in 2018.

Those are just the reported numbers, too. More chips are being made by big systems companies, such as Apple and Google, which are not reflected in the total semiconductor sales numbers because systems companies don’t break them out. On top of that, the SIA predicts that sensors will grow 22.6%, which points to a significant increase in new markets such as IoT, robotics, virtual/augmented reality, and the autonomous and assisted driving portions of automotive.

That’s all good news. The bad news is that several factors could conspire to make it much harder to bring chips to market on time in the future. Among them:

1. Safety-critical design. In the automotive market, and increasingly in markets that intersect in some way with automotive, medical and industrial, reliability requirements are rising. Reliability is a measure of functionality over time, and it requires more than just random fault injection or more thorough verification and debug—all of which will become even more essential. It also demands constant review of data about failures, tracing issues back to the source and fixing them so they don’t continue to cause problems.

Systems vendors have been relying on software, particularly in the computing and mobility markets, to patch problems as they arise. The strategy has been to send out an update, then fix the problems in the next rev of the product. In safety-critical markets, that isn’t always possible, which puts greater emphasis on shifting verification and debug even further left and extending it to the right. This helps explain why the automotive design cycle historically has been two to three times slower than mobile phones. The auto industry is in the midst of a race to put autonomous vehicles on the road, but it’s questionable whether that industry will continue the pace of electronics development once those vehicles go mainstream. The likelihood is that will instead focus on getting the kinks out of these systems and avoiding expensive recalls rather than rushing to the next process node or on-board supercomputer architecture.

2. Scaling headaches become migraines. Chip design is getting harder at advanced nodes because there is more to integrate and there are more physical effects to worry about. At 5nm, the semiconductor industry may be looking at vertical nanowires and a whole slew of manufacturing and material issues, including quantum effects. It takes longer to design and verify those chips, and longer to manufacture them, two pillars in the economics of chip scaling. At that point, the race will likely shift up and out, using new packaging approaches and architectures, a trend that already is beginning to ramp with fan-out wafer-level packaging and 2.5D implementations.

Whether that will lead to a market for chiplets and hard IP, or whether it will lead to platforms that are 80% or 90% complete, remains to be seen. But the basic premise is the same—when there are too many elements and possibilities, even partitioning for a divide-and-conquer approach becomes a monumental task. Making it all work—functionally, electrically, thermally—is even tougher within a fixed market window, unless there are some fundamental changes.

3. Inertia will become more challenging. Methodologies, flows, and silos have taken years to build up and perfect. Many of them will need to be rethought and reconfigured. This is hard enough when there is a roadmap for how the semiconductor industry will change. Roadmaps no longer exist, and while some are being developed for specific applications, it will take time for companies to adjust. Add to that the potential for geopolitical disruption of supply chains, and things could get even more difficult.

Put all of these pieces together and it points to significant changes ahead. The entire semiconductor ecosystem will need to adjust to demands of new end markets, changing economics, and the laws of physics. But once those changes are implemented, there appears to be a whole spectrum of new opportunities that have never been tapped before. And hopefully there will be far fewer questions about when a design will really be done.