News Intel has processed 30,000 wafers with High-NA EUV chipmaking tool

[Admin]

Intel says ASML's High-NA EUV tools have produced 30,000 wafers in a single quarter.

Intel has processed 30,000 wafers with High-NA EUV chipmaking tool : Read more

 

[bit_user]

Processing 30,000 wafers in a quarter is far below what commercial-grade systems can do. However, the number is massive for R&D usage

Can such a number really be just for node development? Or does this imply they're already using it for some low-volume production?

I can sort of see why you might need lots of test wafer for node development. For instance, if you want to sweep different parameter spaces. Then, as the node finalizes, I guess cell library design and testing needs to happen. Still, seems like a lot of wafer to handle, especially if you're doing any sort of detailed analysis of them.

 

[jkflipflop98]

Oh yeah, that's pretty much normal. All the different modules have their own pet projects and issues to iron out. The diffusion guys might be trying a new atomic layer deposition, and the etch guys are trying some new acid/base chemistry for an undercut, while the planar guys are trying out some new pad materials. Then we need fleets of optical inspection tools backed up by hundreds of scanning and transmission electron scopes to double check it all. It adds up quick.

 

[why_wolf]

Can such a number really be just for node development? Or does this imply they're already using it for some low-volume production?

I can sort of see why you might need lots of test wafer for node development. For instance, if you want to sweep different parameter spaces. Then, as the node finalizes, I guess cell library design and testing needs to happen. Still, seems like a lot of wafer to handle, especially if you're doing any sort of detailed analysis of them.

I guess it depends on what a "regular" amount of test wafers is? So this could be a nothing article or it could mean Intel is putting in serious hours on R&D to get back ahead and performing tests back to back.

 

[DS426]

I also had the initial shock of thinking 30K wafers seems high when used just for R&D, but I suppose thousands of unique wafers come about really quickly when several variations exist at many of the different stages. For example, if 4x6x8x4x7 possibilities were used across those 5 stages/characteristics, that would consume 5,375 wafers. Add one more stage with 5 possibilities (different materials, chemicals, time, etc.) and you get pretty close to 30K wafers. :O

 

[bit_user]

I also had the initial shock of thinking 30K wafers seems high when used just for R&D, but I suppose thousands of unique wafers come about really quickly when several variations exist at many of the different stages. For example, if 4x6x8x4x7 possibilities were used across those 5 stages/characteristics, that would consume 5,375 wafers.

Yeah, but I'm sure they wouldn't do uniformly sampled parameter sweeps. Probably something more like a binary search or gradient descent.

If the materials & processing costs are like $10k each, that'd add up to $300M, which seems not way out of line with what I imagine node development costs are running. If that many were fabbed over a single year, it'd be about 82 per day or 5 per hour (assuming equipment is available 67% of the time).

 

[thestryker]

Can such a number really be just for node development? Or does this imply they're already using it for some low-volume production?

I can sort of see why you might need lots of test wafer for node development. For instance, if you want to sweep different parameter spaces. Then, as the node finalizes, I guess cell library design and testing needs to happen. Still, seems like a lot of wafer to handle, especially if you're doing any sort of detailed analysis of them.

Keep in mind the first one completed installation in Apr 2024 and the second in Oct so that's a lot of time relatively speaking. Given that 18A was also originally supposed to be a High-NA node I wouldn't be surprised if they were able to hit the ground running processing wafers quickly. I can only imagine the process engineers salivating at the possibility of seeing what they can do with the node on better machines. I can't imagine if the only thing they were doing was 14A prep this many would have been run through yet.

Depending on the speed with which ASML can get the 5200 to market I could even see one of the advanced 18A nodes being run High-NA since 14A is still a ways out.

edit: Just read the source article and they claim the 30000 is in a single quarter which does seem like a high number, but perhaps that's just an indicator of how rapid development progress is. I imagine we'll find out more in Apr when they do their foundry event.

 

[dalek1234]

We need an insider to tell us what the yield rate is.

If these are wafers for R&D only, it could also mean that there is a high defect rate, and Intel is trying frantically different things to try to improve the yield, hence the high wafer count.

Or it could be a bunch of different reasons, like others have stated. Basically, the article doesn't have enough detail to be of any use.

 

[bit_user]

We need an insider to tell us what the yield rate is.

This isn't for 18A, but maybe 14A is far enough along that they're already starting to focus on yield improvements.

If these are wafers for R&D only, it could also mean that there is a high defect rate, and Intel is trying frantically different things to try to improve the yield, hence the high wafer count.

The article says this was an announcement Intel made at an industry conference. If it were a bad sign, I think they wouldn't have told that to a bunch of experts who'd be best-placed to understand that.

Or it could be a bunch of different reasons, like others have stated. Basically, the article doesn't have enough detail to be of any use.

I think Intel is trying to tell us their development is proceeding at a rapid pace. I agree that it's pretty useless, at least to you and me.

I'm mostly impressed at the processing rate, which as @thestryker pointed out, was 30k per 3 months = 329 per day. If it were going 24/7, that's 13.7 per hour which is almost like printing paper. Even with 2 machines going, that's a lot faster than I thought it'd be.