The number of chipmakers that truly can differentiate their products by moving to the next process node is falling, and that pool will continue to shrink even further over the next few years.
Processor companies such as Intel and IBM always will benefit from scaling and architectural changes. So will GPU companies such as Nvidia, and FPGA vendors such as Xilinx, Microsemi and Altera (now part of Intel), which need a perpetual increase in density. But those are processing engines, and they’re as standard as a USB or an ARM processor core. Other than choosing the best one(s) for a particular job, it’s generally not in the best interest of most companies to develop those.
While finFETs at 16/14nm were a big improvement in controlling leakage current, that benefit begins fading at 10nm and 7nm as leakage returns—along with an increase in dynamic power due to higher density. Even improvements in performance are more difficult because wires don’t shrink. And it gets more time-consuming to develop everything on a single die, which in turn raises the cost per transistor, which was the basis of the formula for shrinking in the first place.
Nonetheless, there is certainly a benefit to having a 7nm processing element connected to a 90nm analog sensor and a 20nm DRAM using well-tested software, and companies are beginning to untether themselves from a shrink-everything mentality and do everything yourself. IP vendors say demand is growing to harden soft IP into chips that can serve as platforms and subsystems, which then can be sold across a wide variety of applications and packaged in multiple different ways.
The big foundries and OSATs are preparing for this shift, which is why they’re building up their advanced packaging capabilities. OSATs are buying other OSATs to gain scale and compete. And chipmakers such as Marvell, as well as design services companies such as eSilicon and Indie Semiconductor, are providing menu options for companies looking to quickly assemble chips for markets.
The word that is popping up repeatedly these days is solutions, not SoCs. This word has been grossly over-used for decades, mostly in the IT and dot-com space, but it has rarely been associated with semiconductors. Still, as companies begin looking beyond standard parts to their own competitive twist on things, that approach will become much more prevalent. It’s no longer about the standard pieces. It’s about the architecture, the microarchitecture, the the package, the interconnects, the software, power consumption, along with some new benchmarks such as reliability and security.
This is hardly bad news for the semiconductor industry. Commoditization of standard parts has been sucking the value out of the chip industry for years, which is the underlying economic formula for Moore’s Law. It’s easy to pit several providers of chips, or even development tools, together if they’re offering comparable products. But it’s much harder to do that with a market-specific solution that cannot be measured against another standard part. That means average selling prices will rise across the board, and the value of getting those to market more quickly will provide a premium for those who can deliver different architectures, packaging approaches, and tools that shorten the design-through-manufacturing cycle.
What’s clear is that the value in semiconductors is shifting. The big question now is who will benefit most from this shift—and who won’t.