The IEEE’s plan to add structure for individual markets is an important step.
The semiconductor industry is now officially looking beyond PCs and servers, establishing metrics and guidance for existing and developing market segments rather than just focusing on how to get to the next process nodes.
The IEEE’s International Roadmap for Devices and Systems marks a fundamental shift in the industry. The uncertainty that has ensued ever since the introduction of 3D transistor structures and multi-patterning has forced many companies to stop and consider what comes next and how much they’re willing to gamble on those opportunities. In some cases, the price of entry can be huge, and the return uncertain. The slowdown in smart phone sales means that no single chip will sell in the billions of units. And while everything increasingly is connected, most of those connected systems will not generate the same kinds of volume.
But the trouble runs deeper and wider than that. Processor architectures that have been developed for computing or inside of smart phones are not the best way to do pattern recognition or search or many other tasks for which they are now used. In some cases, it’s like using a shotgun to kill a fly. It may be better than a cannon, but it’s still not the best choice.
Nowhere is this more apparent than in battery-powered devices. Too many cycles, too much data being pushed in and out of memory, and trying to manage fluctuations in compute jobs by throwing more homogeneous cores at a problem isn’t an ideal solution. There’s a reason why many people don’t strap on their wearable devices every day. The battery isn’t charged.
For the IEEE to get to this point, with strong support across the industry, means that something is broken. It’s not broken everywhere. We’ve said for some time that Moore’s Law isn’t dead. But it has become less relevant for many chipmakers across many markets. It doesn’t matter if a smart refrigerator has eight cores or whether an embedded sensor can communicate with memory at gigabit speeds. What matters is that designs are correct for their specific application, and that chipmakers have some guidance for what’s needed to enter those markets.
This is a very big deal. It explains why there is experimentation with new packaging types—2.5D, 2.1D (organic interposer), 3D IC, Marvell’s MoChi, TSMC’s InFO and CoWos, Intel’s EMIB—as well as more options at established nodes, including low power and high performance versions stretching back to 90nm, as well as FD-SOI at 28nm and 22nm. There are new memory types in the works (Magnetoresistive RAM, ReRAM, 3D-XPoint) new memory architectures (HBM, HMC, DRAM modules on server racks), and new ways of connecting all of this together.
These solutions are being developed in isolation, sometimes by company, sometimes by industry segment. Bringing them together under one roof with a roadmap for what’s different and changing in each market sector is a big step in the right direction. Execution will be critical, of course, and the real challenges are still ahead. It’s one thing to pull together a single roadmap for an industry. It’s quite another to create effective roadmaps for a number of sectors that feed that industry.
But the upside, if done right, is huge. There will be many billions of things connected over the next decade, and most of them will be purpose-built and quite complex in their own right. Done correctly, this could prove to be the biggest opportunity in the history of semiconductors, from tooling on the front end to manufacturing on the back end. It could expand the focus from chips to systems, and open the doors to many new markets that never existed in the past.
The IRDS is a first step in that direction, and its importance should not be underestimated.
Rethinking Processor Architectures
General-purpose metrics no longer apply as semiconductor industry makes a fundamental shift toward application-specific solutions.