Lithography in a quantum world

[Fred Chen]

Another very interesting article by Harry Levinson.

ShieldSquare Captcha

Good to have another voice speaking to these issues.

 

[rchang94]

Thank you for sharing! Do you think tools like Sculpta would be the key in addressing the challenges with LER?

 

[Fred Chen]

Thank you for sharing! Do you think tools like Sculpta would be the key in addressing the challenges with LER?

Etch is often a tool used to mitigate some litho-origin problems like roughness. But etch cannot overcome problems of stochastic edge placement. The feature (relative) center of mass cannot be moved by etch.

 

[rchang94]

Etch is often a tool used to mitigate some litho-origin problems like roughness. But etch cannot overcome problems of stochastic edge placement. The feature (relative) center of mass cannot be moved by etch.

Thank you. I was thinking about the roadmap challenges shown in the paper with LER and LWR at below N2 and etch might be an alternative way to solve this vs. relying just on resist development. Also I wonder what your thoughts are on the pricing of EUV tools given he mentioned it towards the end but didn't go into it given its outside the scope of this paper.

 

[Fred Chen]

Thank you. I was thinking about the roadmap challenges shown in the paper with LER and LWR at below N2 and etch might be an alternative way to solve this vs. relying just on resist development. Also I wonder what your thoughts are on the pricing of EUV tools given he mentioned it towards the end but didn't go into it given its outside the scope of this paper.

The EUV systems are bigger than DUV because the mirror optics cannot be collimated, and the larger angle range of High-NA makes those systems yet bigger. The going price must be at least 200 million.

 

[rchang94]

The EUV systems are bigger than DUV because the mirror optics cannot be collimated, and the larger angle range of High-NA makes those systems yet bigger. The going price must be at least 200 million.

I've heard high-NA in the range of 300-350m. Just sounds like a crazy expensive tools which based on reading a lot of your posts have some pretty significant flaws. However ASML CTO in interviews seems very confident on its adoption in HVM. I'm wondering if I'm missing something here or could there be a material risk that high NA flops.

 

[Fred Chen]

I've heard high-NA in the range of 300-350m. Just sounds like a crazy expensive tools which based on reading a lot of your posts have some pretty significant flaws. However ASML CTO in interviews seems very confident on its adoption in HVM. I'm wondering if I'm missing something here or could there be a material risk that high NA flops.

They already have some customers presumably for the R&D purpose. The resist response (defocus, electrons, collapse) can be prohibitive for the high-NA EUV, especially talking about <20 nm thickness.

 

[Tanj]

They already have some customers presumably for the R&D purpose. The resist response (defocus, electrons, collapse) can be prohibitive for the high-NA EUV, especially talking about <20 nm thickness.

Which is why the R&D is going on, to refine the production recipes.

A key part of the EUV roadmap is ramping up the illumination to get better throughput. The ASML view is no so much that customers should go from multi-exposure ArFi to single-exposure EUV. Yes, customers do, but that is because throughput is low. But as ASML ramps up throughput the customers will use double or even triple masks in EUV 0.33, and they will transition to multiple exposures at 0.55 as well, for a natural increase in capability. Sure, in some use cases they will want to run a single exposure. But ASML is investing a lot of the $300M into throughput, automation, and uptime, not just optics, to ensure that customers have good economics no matter how many exposures are combined for a pattern.

Yes, 0.55 will have quirks, but if you learn the right recipes and use the machines correctly, they will deliver.

 

[Tanj]

I've heard high-NA in the range of 300-350m. Just sounds like a crazy expensive tools which based on reading a lot of your posts have some pretty significant flaws. However ASML CTO in interviews seems very confident on its adoption in HVM. I'm wondering if I'm missing something here or could there be a material risk that high NA flops.

Lets assume the average pattern is a dual exposure when using 0.55NA. Just to test the economics. The machines are highly automated, so you run maybe a 100 wafers through the first mask, swap masks and adjust pupil and aperture, then run them through the second mask. All automated and finished in under an hour, so the shelf life of the resist is acceptable. It helps that the resists are probably applied in a preload chamber of the machine, and the resist is fixed, before exit.

So, if you get 8000 hours of up time and average 100 WPH that is 800,000 wafers through the machine. Allocate a $300 value to the layer - which is high but not extreme - and you minted $240M value. Yes there are maintenance costs but you can also argue there are costs saved by having more capable layers and fewer of them. Also, the chassis is upgradeable so you can plan to use that beast for 5 years on leading edge chips, maybe spending $500M or more including the upgrades, but still easily earn the purchase cost.

 

[Tanj]

There is another factor in the cost. You may remember when Intel was unchallenged for CPUs you might have to pay 50% extra for a CPU with 20% higher clock rate. That did not translate into even 20% more throughput - but the chips sold because they multiplied the value of your whole system, not just of the CPU. So Intel was sucking in the margins for the whole PC.

In the same way if 25% of your foundry is equipment allowing you to make leading edge chips keeping the other 75% of your CAPEX busy on the highest margin work, then your 25% has leverage to be worth much more than just its own productivity.

 

[Fred Chen]

The ASML view is no so much that customers should go from multi-exposure ArFi to single-exposure EUV.

This was their promoted view, it was just wrong. 5nm came in heavily multi-patterned.

The quantum perspective means "EUV" is not EUV but low energy electrons as well, driving the resolution and reliability.

 

[nghanayem]

This was their promoted view, it was just wrong. 5nm came in heavily multi-patterned.

How do you mean Fred? Just M0 and fins were MP. All other metal layers are well into the direct print euv range or DUV SADP range. On the ASML front that was definitely the promise, but I don’t think it was the eternal promise. Stochastic defects or no, you would eventually have to go to MP if you don’t go to a more advanced scanner tech. Given how EUV SE was replacing LE^3 and early SAQP, it makes sense that era was shorter than if EUV was ready on time and replaced dry 193nm. I guess I would call it ~1.5 TSMC node gens of direct print. Probably a bit shorter than what would have been ideal, but I don’t think that this is as big of a bait and switch as I think you are making it out to be.

 

[Fred Chen]

How do you mean Fred? Just M0 and fins were MP. All other metal layers are well into the direct print euv range or DUV SADP range. On the ASML front that was definitely the promise, but I don’t think it was the eternal promise. Stochastic defects or no, you would eventually have to go to MP if you don’t go to a more advanced scanner tech. Given how EUV SE was replacing LE^3 and early SAQP, it makes sense that era was shorter than if EUV was ready on time and replaced dry 193nm. I guess I would call it ~1.5 TSMC node gens of direct print. Probably a bit shorter than what would have been ideal, but I don’t think that this is as big of a bait and switch as I think you are making it out to be.

Samsung showed stochastic defects at the 36 nm pitch: https://semiwiki.com/forum/index.php?threads/2022-euv-stochastics-status.16325/ So at least the initial rollout (which had 36 nm pitch or less) should not have been SE.

 

[Tanj]

This was their promoted view, it was just wrong. 5nm came in heavily multi-patterned.

The quantum perspective means "EUV" is not EUV but low energy electrons as well, driving the resolution and reliability.

I was using present tense. You are right, that was how the expectation built up before EUV was in production, but as EUV was slow to arrive the DUV practice kept getting better and it made EUV SE less competitive. Stochastic issues too. ASML get plenty of customer feedback, they have pivoted to building (and upgrading) machines which will be profitable to own for multiple exposure and higher doses on each exposure.

Yes, low energy electrons will need to deliver about 20 electrons per nm2 (300 uC / cm2) to have stochastics which can deliver resolution competitive with EUV.