The 7nm Pileup

The number of 7nm designs is exploding. Cadence alone reports 80 new 7nm chips under design. So why now, and what does this all mean?

First of all, 7nm appears to be the next 28nm. It’s a major node, and it intersects with a number of broad trends that are happening across the industry, all of which involve AI in one way or another. The big question now is how many of them will survive long enough to move to 5nm or 3nm.

There are literally scores of chip startups vying for a piece of this market, and it’s evident in the amount of funding available to hardware companies for the first time in decades. Several years ago, the number of hardware startups could be counted on one hand. These days, you need a lot more hands. VCs are literally throwing money at these companies, which means they see a strong likelihood that some of these startups will hit big, while others will be acquired.

Initially, much of this computing was contained in the large data centers. But as clouds begin drifting out of the data centers toward the edge, however fuzzy that definition, there is a massive new opportunity emerging. Even more important, there is no clear winner at this point. In fact, as Steven Woo, Rambus fellow and distinguished inventor,  aptly observed, “there is no incumbent instruction set in the edge.” In plain terms, no one owns this market yet—not Intel, Arm, RISC-V, MIPS, or any other startup with a better idea. And nothing pulls in investment dollars/Euros/renminbi like a greenfield opportunity in semiconductors.

Not all of these companies will survive, of course, but the discontinuities here are significant. By using a combination of new high-throughput heterogeneous architectures, 7nm transistor density, and advanced packaging connectivity to high-speed DRAM, performance can be increased by at least two orders of magnitude, and possibly much more. Moore’s Law provides the density, but the big performance and power improvements are well outside of what shrinking features provides.

There are other forces driving this node, as well. Automotive companies are on a 4- to 5-year development cycle, which may seem slow until you consider they used to have a 7- to 10-year cycle. There is a lot of detail to digest in automotive design, and a system developed today—particularly the AI brain in an autonomous vehicle—will be extremely dated in 5 years if it’s developed at 40nm. If people hang onto vehicles for an average of 10 to 12 years, that will be like trying to run complex algorithms today using a Pentium 90.

What will sell cars in the future won’t be the rumble of the engine under the hood. It will be the electronics in and around the car. A 7nm design at least will carry carmakers into the early part of the next decade, if not beyond.

The adjoining piece to chip performance is the throughput of data outside of a device. That is being fueled by massive numbers of sensors everywhere and opportunity to spot and leverage patterns within that data. Most of the network switch and infrastructure has been working at 28nm using custom designs. But Andy Bechtolsheim, founder of Arista Networks (and Sun Microsystems), said at CDNLive this week that the chips in switches will be migrating to 7nm over the next year or two, which will significantly improve performance due to much greater density. Those are more generic chips, too, which means they will be less expensive than building a new chip from scratch (estimated to be somewhere in the neighborhood of $300 million at 7nm), and there will be an awful lot of them sold.

The big question now is what comes next. Will companies spring for $500-plus million for a 5nm chip? And will they be able to extend architectural gains beyond the current round, or will this be a one-time gain? And will 7nm be the really next big node, where the improvements in performance, power and cost are less about process than other factors. No matter what, this is a much different market than anyone would have predicted 18 months ago.

Related Stories
New Design Approaches At 7/5nm
Smaller features and AI are creating system-level issues, but traditional ways of solving these problems don’t always work.
Power Delivery Affecting Performance At 7nm
Slowdown due to impact on timing, and dependencies between power, thermal and timing that may not be caught by signoff tools.
Variation Issues Grow Wider And Deeper
New sources, safety-critical applications and tighter tolerances raise new questions both inside and outside the fab.