For fifty years, Moore’s Law has relied on advances in lithography to propel the semiconductor industry forward. Optical lithography has achieved a 250x improvement in resolution over that time, all while keeping the cost to print a square centimetre of silicon about the same. But current lithography tools have been stuck at a minimum resolution of 40 nm for the last eight years, necessitating expensive multiple patterning to break this limit. As a result, patterning costs are rising faster than they ever have, putting the continuation of Moore’s Law in doubt.
Extreme ultraviolet lithography was supposed to fix this. Its lower wavelength enables sub-20-nm resolution, beyond the capabilities of double patterning with 193-nm immersion lithography. But resolution is not enough to save Moore’s Law – it must be cost-effective as well, while providing the same or higher manufacturing yields. And this is where EUV continues to fall short.
There are three technical roadblocks to a successful introduction of EUV lithography into manufacturing, and each of these plays into its economic potential. The first is mask defectivity. A defect on a wafer can kill one chip, but a defect on the mask can kill every chip. So every mask must be made and kept defect free. We don’t know how to do this yet for EUV masks. Mask blanks, the starting material for making masks, have too many defects. We’re not good at finding every defect that might exist on a mask, and not all of the defects we find can be repaired. And once a mask is made, we don’t yet have a working pellicle (a transparent wrapper that keeps defects away from the mask while in use) to ensure the mask stays defect free.
The next problem is well known: source power. Generating bright EUV light is very, very hard. But if the source is too dim, the tool’s throughput suffers, ruining its economic benefit. Opinions vary, but many people think an EUV source output of 250 W will make EUV lithography economical for production. I think a factor of 2 to 4 higher than this will be needed to achieve sufficient manufacturing yields due to the final technical roadblock for EUV: roughness.
When printing in photoresist using low power, the resulting patterns become very rough due to fundamental statistical limits (called stochastic effects). Today’s best results are too rough by a factor of 2, and to smooth them will take an increase in exposure dose of something like a factor of 4. But unless the source brightness is also increased by a factor of 4, the result will be an unacceptable drop in throughput.
Is this problem solvable by means other than higher source power? Many people hope that some magical photoresist formulation will come to the rescue, but resists work under the same rules of physics as do lithography tools. I’m sceptical that new chemistry – which can only add noise, not take it away – will overcome these statistical limits.
Which brings us to the future of EUV lithography. Its proponents have been fond of saying that EUV is now a matter of when, not if. But such sentiments ignore the reality of life under Moore’s Law: a matter of when is a matter of if. Solutions that arrive too late are never successful, because the industry simply moves on without them. We needed a working EUV lithography process at the 22-nm node, and at the 14-nm node, and at the soon-to-be-released 10-nm node. But in the absence of a manufacturing-ready EUV tool, the industry found other solutions. And other solutions are being developed for the 7-nm node as well, since semiconductor manufacturers are unwilling to bet their future on unproven EUV.
So if EUV is too late, it will run out of capability before it has a chance to be used in manufacturing. We’ve made great progress on all of its technical roadblocks, but much more progress is still needed. And time is running out for EUV lithography.