Let us first apologize because two of the three Common Platform partners, IBM and Samsung, forbade taking pictures and did not give out slides. Global Foundries was fine with pictures, but they sadly did not give any of the EUV talks. SemiAccurate has to question the logic of forbidding attendees and press from taking pictures when the entire conference is webcast to anyone who want to sign up. Competitive intelligence folk don’t even need to put on pants to get all the info, but legitimate attendees can’t post slides or pictures. *SIGH* That means lots of words, not blurry shots like we normally do.
Getting back to EUV, getting EUV ready is a problem that progress is best described as glacial. EUV stands for Extreme Ultra-Violet, specifically a 13.5nm wavelength light source and attendant photo-lithography tools. This replaces the tried and true 193nm light sources currently in use by the industry for many years, it is a known quantity, widely trusted, and completely stretched to the limits of it’s capabilities and beyond. With 193nm light, to get the line widths needed, expensive and yield lowering double and triple patterning is needed. With EUV it is one exposure, but everything is new and different.
EUV will allow foundries to etch lines far smaller than those possible with 193nm light sources, but EUV is not without new and exciting problems. By exciting, most people mean big and very hard to solve, sometimes even worst than that. The first minor catch is the light source itself, 13.5nm is not even close to visible light. In fact, EUV is the polite term for soft X-Rays, and that slice of spectrum is extremely difficult to generate light at. Most current sources are basically a chunk of radioactive material in a shielded box. You open a door and X-Rays come out. It works, but is both far too imprecise for chip fabrication work, not to mention being the wrong wavelength.
The mechanism to solve this should be quite obvious to anyone who has opened an encyclopedia sometime in their life. You start out with microscopic drops of molten tin dropped in to a hard vacuum chamber. It has to be a hard vacuum because anything less will absorb and possibly diffuse the EUV light. When dealing with lines that are a few tens of nm wide, it doesn’t take much diffraction to ruin a good pattern. You also need the light to be very precisely placed. This is why you need to shape those droplets of tin in to a disc instead of a droplet, so you hit it with a low power laser in mid-air, err mid-vacuum, to physically change the shape. Once in the correct shape, you must again hit the same drop with a very high powered laser to vaporize the tin. This releases EUV light of the correct wavelength, and if you shaped the droplet right, in the right pattern.
This light is shot back in the direction of the laser, and luckily away from the vaporized tin fragments that have just been blasted by an extremely powerful laser pulse. That perfectly formed 13.5nm EUV light pulse, being an X-Ray, can not be used with transmissive lenses, you need absurdly accurate reflective lenses to focus the EUV pulse and point it at your target. That target is also likely in the direction you just shot molten tin at a significant fraction of the speed of light at. Since the EUV beam is really sensitive to perturbation by things like atoms of gas, a pellicle to protect your shatteringly expensive mask set is right out, sorry.
Assuming this all works, and you don’t end up with tin flecks embedded in your multi-million dollar mask set, congratulations, you have just made one exposure for one mask layer in your EUV chip. Repeat this a few hundred times per wafer, and voila, you have a controller for the next Furby or something. Wasn’t that all worth it? [Editor’s note: If you cannot follow the bouncy-light or should we say x-ray path, this link has a picture that might help visualize the process.]
That, however, is the easy part. The first currently unsolved problem is the light source. Current technology can provide EUV light with a sustained output of about 30W, roughly 1/3rd the power of a decent incandescent bulb. According to IBM, current developments will get us to 125W or so, a large jump from where we are. Sadly, 250W sustained is needed for actual production from these room sized behemoths. If 250W doesn’t seem like much, that is the energy output, input is orders of magnitude higher.
In the end, if you get a small fraction of the input power out as EUV light, say high single digit percentages, you should be overjoyed. What, did you think maintaining a hard vacuum, melting tin, physically shaping it with a laser, then vaporizing it with enough energy to create X-Rays was doable with 110v wall power? Tens of thousands of watts input, 30W out with the hope of 250W someday. Calling the process energy intensive and inefficient is almost trivializing the problem.
From there, the next big problem is masks. Masks are multi-layer beasts that are anything but simple. Think about it, you are masking X-Rays to draw lines in the low tens of nm wide, not something you can make out back in the shed. Not only that, but the mask blanks have to be perfect, defects in the single digit nm range can mean a blank is scrap, and they cost hundreds of thousands of dollars each. Last year there was no real way to inspect blanks, but that seems to have been overcome in the intervening time, at least inspection is not a one by one manual job any more. That is big progress.
To make matters even better, defects in mask blanks can now be repaired. Unfortunately, that is only if the defect is in the top layer. If it is in a lower layer, and you are lucky enough to be able to detect it prior to making the mask, well tough, some executive just got a paperweight that is worth more than her car. One way of dealing with the problem, if detected while still a blank, is pattern shifting. This means you know where the defect is, and you move the pattern over a few nm so that the defect ends up in a spot that does not cause problems to the pattern being drawn. If you can do that, you are set and defects in underlying layers are ignorable. If you have two defects, or it is a big divot, oh say 20nm across, see the paperweight comment above.
What it all comes down to is that EUV is a really hard technology to develop. Progress is being made with masks moving forward faster than other areas in the past year or two. Light sources are still problematic, power inputs sufficient to meet the needs of a small neighborhood only produce outputs of a weak light bulb, but the players have hope to match a bright light bulb soon. We won’t point out that this is the same, “Chin up, pip pip” attitude we have been hearing from the EUV camp for half a decade now, but who are we to criticize. Oh wait, that actually is our job. Still, they have it tough, and we don’t doubt that things are moving forward here.
In the end, EUV was going to be ready for the 14nm node a few years ago. More recently, that intercept moved to the 10nm node, and now it could still be ready at 10nm, but most foundries are planning for it potentially not happening until 7nm. If EUV is ready before 7nm, there are plans to slipstream it, but no one SemiAccurate asked was willing to count on EUV at 10nm. And that is where things lie, two steps forward, one step backward, the dance currently is slated to end at 7nm. Maybe. More next year.S|A
Latest posts by Charlie Demerjian (see all)
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018
- Why Can’t Intel Supply Enough 14nm Xeons? - Sep 10, 2018
- Intel can’t supply 14nm Xeons, HPE directly recommends AMD Epyc - Sep 7, 2018
- AMD reintroduces the Athlon name with two CPUs - Sep 6, 2018