Hacker News Reader

Comment by curious_cat_163 | Hacker News Reader

Would you share some insight on why Intel failed at making EUV work? Gelsinger said on a podcast (Decoder) recently that in fact they fought it.

Obviously, they are kind of big enough to do both — make a half hearted attempt and fight it all the same due to some internal politics. However, for something this existential, one would think that the smart people and the leadership would try to put some focus on the field. Was it that bad? Really, looking for some insight on what really went down.

View on HN

Intel (like every other chipmaker) has been chasing EUV for decades, but every time they thought they finally nailed it they turned out to be wrong.

The key that ultimately made it possible was the light source, and ASML is the only company that sells these light sources. ASML didn't invent the light source entirely on its own; there were a lot of ideas for light sources floating around and the only one that turned out to work was laser-induced plasma from tin microdroplets, which sounds completely crazy but it works [0]. ASML perfected that process (with -- to be fair -- some investment from Intel).

[0] https://www.asml.com/en/technology/lithography-principles/li...

Intel did fight EUV with a technology called SAQP which ultimately turned out not to work as well as EUV, but they were researching EUV at the same time. Here are some more historical articles about Intel's work on EUV in the past. Note the dates.

https://newsroom.intel.com/news-releases/intel-achieves-majo...

https://semiengineering.com/why-euv-is-so-difficult/

https://www.eetimes.com/intel-developing-euv-pellicle/

https://www.extremetech.com/computing/276376-intel-reportedl...

Do you have any insight on what organizational failures led to them failing to find the right EUV solution?

Was it just a matter of luck that ASML nailed the right one or was it the nature of exploring some riskier alternative (as you said "which sounds completely crazy but it works")?

I know nothing about Intel's internal organization.

There were several different light source ideas tried. Some at US national laboratories. Tin plasma was known to work but it left so much debris that it gummed up the optics so it was forgotten for a while. The key turned out to be tin microdroplets.

https://www.optica-opn.org/home/articles/volume_29/march_201...

ASML and Intel are two completely different companies. ASML makes lithography machines which is one of the many steps in producing a chip. ASML machines are used by all chip manufacturers. The challenge for chip manufacturers like Intel, but also tsmc, Samsung among others, is to optimize the entire production process. And that has been / is not easy with a lithography step (EUV) which is so drastically different then what the industry has been using before. One of the challenging choices for chip manufacturers was to drastically change the process and implement EUV or to go self-aligned quadratic patterning. The latter, if you take a simplistic view, can be seen as an evolution on existing processes. We know see that those companies who went aggressively into EUV are having an edge on their competitors.

When they say “microdroplets” (I may be out of date here) they mean firing lasers at falling drops of molten tin and then capturing the energy that the tin gives off.

Like anything, some of their PHDs were like “no that’s probably not going to work because xyz” and they trusted said PHDs.

It’s just science. It’s hard to predict.

I think it's more that they didn't try ('till now) 'cause they thought ASML's machines just weren't ready for prime time. TSMC bought ASML's bleeding edge EUV machines thinking they could make it work eventually, and they did. Intel may catch up very quickly now.

I thought that ASML got its EUV light source, basically, by buying Cymer in San Diego. Read some of [0], and then searched it, but didn't find a single instance of 'Cymer.'

Thank you.

I am much more hopeful about Intel's chance at EUV than GP. In order to avoid repeating myself, here's my analysis of what went wrong: https://news.ycombinator.com/item?id=32335873

Why is it that it's taking them so long to pivoting from that failure? 10nm didn't work out, that's fine. But it feels like that didn't work out and they've spent the years after that in a hole and still haven't dug themselves out.

From the rumors I've heard (might be true; might not) Intel lost a great deal of semiconductor engineering talent because they were making so much money on minor evolutions of x86 that they saw little need to truly innovate. The MBAs took over.

Once it became clear that AMD and ARM (and by extension TSMC) were eating their lunch, they had great difficulty pivoting because they lacked the talent and the organizational will to do so -- at least until Gelsinger took charge. I have a lot of respect for Gelsinger and I think he just might turn it around.

> ...they were making so much money on minor evolutions of x86 that they saw little need to truly innovate. The MBAs took over.

Yet another example in a long history that modeling your talent pipeline as a JIT supply chain does not work in high cognitive input settings. This is a general unsolved management challenge: capturing sufficiently granular and comprehensive institutional knowledge to be able to efficiently restart organizations that rely upon it years down the line.

I don't think we will ever solve it until we learn how to share human mind states. The complexity of these high cognitive input activities at large scale has surpassed the ability of our capabilities to transfer comprehension between minds using conventional information transfer techniques (primarily written word and audio video). Even were I to guarantee the capture of all information at sufficient granularity, the only way to QA that is to de novo staff up a brand-new parallel organization, and have them reproduce the original organization's accomplishments. No one is ever funded for that.

This is a really interesting problem space for me because I suspect it is a dimension of technical debt very rarely explored because our species has not had to really grapple with it before as our technological progression has been slow enough we could "good enough" brute force our way through it whenever it came up in less severe forms.

> Would you share some insight on why Intel failed at making EUV work? Gelsinger said on a podcast (Decoder) recently that in fact they fought it.

A cursory Google search results in multiple search hits from recent news pieces announcing Intel buying their EUV kits from ASML.

If Intel is handing out small fortunes to ASML to deliver a machine that in the past Intel was expected to develop in-house, that does not sound like a success story.