A (0:03)

All indications point to TSMC's N2 process node being a beautiful one. TSMC recently discussed their 2nm process node at IDM 2024, calling it the world's most advanced logic technology. N2's headlining feature is of course the Gate all around transistor, its first major transistor transition in nearly 15 years. But there is another new technology being inserted into the node, the curvilinear mask or masks with curved lines. TSMC's first node with curvy masks curves. Big whoop, right? It matters because these masks unlock the power of GPUs for semiconductor manufacturing. In this video, TSMC's curvy masks for its 2nm node photolithography uses light to shrink and project a chip design image from a special glass or mirror template called a photomask to a wafer. After exposing the image, we develop it and then do etch processes to solidify it. The photomasks themselves are prepared in special facilities called mask shops. Inside, special machines called mask writers. Write the chip pattern onto 6 inch mask blanks using special tools firing electron beams. Because the masks will be the template for millions of chips, they must be perfect. Mask features are larger than on the wafer, usually four times larger because of the optical magnification. But defects will print on production wafers, potentially costing millions or billions of dollars. TSMC detailed in a paper how the accuracy of their mask shops in Taiwan can be affected by natural disasters like earthquakes. Yes, that one obviously. But also typhoons because they affect atmospheric pressure inside the cleanroom. Even shifts in world magnetic fields can destabilize the mask writer tools. I cannot emphasize just how precious these masks are. A complete mask set for a modern integrated circuit might have anywhere from 40 to 70 masks. Each mask takes about 24 to 72 hours to make. And the whole set can cost tens of millions of dollars. Beyond the financial cost, masks contain the Fabless company's IP designs that billion dollar companies like Nvidia, Apple, AMD and so on have invested hundreds of millions of dollars into. It's imperative that these are protected and done right. As I see feature sizes shrink, it gets harder for a photolithography machine to faithfully transfer patterns to the wafer. Light propagates through the machine in certain ways, diffracting and distorting despite engineers greatest efforts. As a result, features might end up too close together. The widths of lines might be off spec, so on. One way to address these issues is to build a machine using a smaller wavelength. A Smaller wavelength gives you more resolution, but engineering that is hard and costs money, and we can't do that for every new node. Fabs in general prefer to switch over only when they absolutely have to. So as an intermediate solution, they adopted what are called resolution enhancement techniques, or rets. RETS are adjustments made to the photomask to improve the overall resolution of the chip design's transferred image. A simple example being to add serif like tick marks to the ends of shapes in the chip design pattern. Such tick marks and others like it called sub resolution assist features are not generally intended to be printed onto the wafer, but having them there improves the resolution of the design. At the start, these began as simple heuristics, but over time they have increased in complexity and effort. Computer models were even employed to rework the photomask design with increasing sophistication, but these tricks were always done by taking the final photomask design and adding things to it. Inverse lithography technology, or ilt, approaches the problem from literally another direction. I discussed ILT in a prior video. ILT first investigates how light travels through the lithography optics and interacts with the photoresist. Then, knowing that the chip pattern is what we want transferred to the wafer, it generates the optical mask image for that pixel by pixel. This backwards direction from wafer to mask is the key differentiator and why the word inverse is in the name taken to the fullest extent. These post ILT designs look alien, almost psychedelic. But these weird designs also work incredibly. They widen the process window, meaning that even if deviations occur in the lithography process, what ends up being printed is more likely to still be within spec. In other words, it gives you flexibility and options. Looking at some of the crazier ILT generated mask images that come out, you might notice that they all have curves or circles. Post ILT mask images almost always have such curves because light physics do not like 90 degree angles. To find out why that's a problem, we need to go to how masks are made inside the Masks shop. They are produced via electron beam using a tool called a variable shape beam or vsb. In a vsb, a single electron beam is steered and focused with electron lenses. One lens inside the machine shapes the beam via an aperture and then a second lens adjusts the beam's size. The big thing about the vsb, however, is that it can only produce rectilinear shapes, meaning straight lined shapes onto a photo, rectangles, squares and 45 degree triangles and of differing sizes. Such features are Said to be aligned to Manhattan geometry X and Y axes like the city blocks of Manhattan. Since the machines cannot produce curved lines, mask makers do a conversion process called Manhattanization. This is where you fracture the curved line, if any, into overlapping rectangles. Manhattanization is not ideal. You suffer a loss in fidelity like a non retina phone screen where you can see the fuzziness in the edges of things. And it requires an additional round of computation after first producing the ILT design. So adding to the turnaround. And finally, a VSB writes each rectilinear shape with a single beam shot. So by replacing a curved line with a whole bunch of blocks, Mahonization adds a bunch of blocks that the VSB must produce, drastically slowing down throughput. End result out. ILT remained an academic concept preached by professors in lecture halls and papers, but never practical enough for fab floors. But things are finally starting to change with two new developments. The first being the rise of a new generation of mask writers wielding hundreds of thousands of electron beams. The multibeam mask writer. How do these new multibeam writers work? Like the vsb, they start with a single beam. But that beam then passes through a special aperture plate system which splits it up into an array of a quarter million or so beamlets, each about 10 to 20 nanometers wide. Small electrodes inside the aperture plate system can electrically deflect its corresponding beamlet towards a stopping plate. The beamlet then disperses. So basically you can think of each beamlet as like a pixel in a monitor screen that can turn on or off. And by fiddling with how long the electrode is open, you can get more precise fine tuning. Like shades of gray in between black and white. After this, a set of projection optics shrink the size of the beam lit array by 200 times to give us the precision we desire. Also in this stage, the machine raises the beam's kinetic energy to about 50,000 electron volts. This is delicate and requires extra engineering, but reduces electron impact damage to the aperture plate system above. Finally, the beamlet array hits the glass or mirror mask plate, referred to as the reticle. The reticle is held in position by a carrier called called a stage. Precise mechatronics move around the reticle just like how an ASML machine positions a wafer. And why did the fabs adopt these multi beam writers in the first place? The original answer was for uv. The two major multibeam mask writer companies are IMS Nanofabrication in Austria and Nuflare in Japan. Let us start with IMS Nanofabrication which I must note, is not the same as the company IMS Fabrication that one produces autoclave tools and equipment for the retreading tire industry. Founded in February 1985 in Vienna, IMS Nanofabrication began as a startup developing a lithography concept called ion projection lithography. Ion projection works by firing electrically charged hydrogen or helium ions at high speed through a carbon coated silicon mask partly funded by Europe's Eureka project. The efforts showed promise but ultimately failed to gain traction. When it became clear in 2001 that the industry would anoint EUV as its next generation lithography system, IMS pivoted to multi electron beam mask writing tools. It was then a risky move. These mask writers present many tricky engineering challenges, identifying and capturing defects, particularly in tricky areas. Streaming the chip design's huge amounts of data to the machine accurately and at high speed. Development will be expensive, and they wouldn't be the only ones trying. Kla Tencor spent $226 million of their own money and DARPA's money on multibeam mask writers before abandoning the project in 2014. IMS partnered with a Japanese electron microscope company called Joel to help with the development. Then in 2012, they demonstrated a proof of concept multibeam mask writer device that put them about three years ahead of the rest of the market. Intel took an interest in IMS's technology and even invested in them in 2009. When things at IMS started to falter in 2015ish, perhaps due to financial reasons stemming from delays, intel stepped in and bought them. With this, IMS had the resources to finally roll out the industry's first commercial multi beam mask writer, the MBM W101. The machine began alpha tool testing in 2016 and hit high volume in 2017. For TSMC and the other fabs, the proximate reason for buying these machines was EUV. EUV's insertion in 2017 was expected to reignite Moore's law, blowing up photomask complexity and sending mask making throughput into the dumps. Though I should note that even before euv, increasingly complex designs were creating too many shots for VSB machines and their single electron beams. Average mask write times had risen from 8 to 9.6 hours in 2015. So all in all, multib mask writers were seen as key to potentially reducing the times of making the most complex masks from 30 or even 72 hours to just 10. The presumption at the time of the 2016 intel acquisition was that Chipzilla would leverage IMS's technology to catapult themselves into the EUV era. But perhaps because of their own money issues, they've been slowly selling stakes in IMS to Bain, a private equity group, as well as tsmc. New Flare is the second major mask writer company in the space. They began in 1976 as a division inside Toshiba Machine, building equipment for the iconic Japanese company's semiconductor business. Through their Toshiba association, they joined the Japanese government's now famous VLSI project. The project helped fund the RD for the variable shaped beam that now makes up the core of their mask riders. In 2002, Toshiba machine spun the division off as its own company, New Flare. The name is made up of two parts, New for the Greek word for new and it also references Pneumazu, the city of their founding Flare stands for their mission to start a new flame. Nuflare started producing mask writers, mask inspection systems and epitaxial growth systems, the first such being their specialty with tools like the EBM series of riders. A month after IMS published their proof of concept multibeam rider in 2012, Nuflare began developing their own tool, the MBM1000. The machine was targeted for release in late 2017 to come out at about the same time as IMS first machine. Nuflare hit that deadline which kinda surprised IMS as they thought they had a multi year lead and since then the two companies have vigorously competed in the mask writing industry, though IMS remains the market leader. As I implied earlier, the original use case for these multibeam mask writers was throughput, getting out these increasingly complicated masks faster. But ILT and the curvy lines had long been the multibeam mask writer's second killer app and after several years TSMC is now finally moving it to high volume production. By carefully moving the reticle stage and turning the pixels on and off, the mask shop can effectively paint the chip design's curved lines onto the reticle. Improvements had to be made to the multibeam mask writer's mechatronics and diagnostics capability and they had to get more fine tuned control of the individual electrodes inside the aperture plate system. But perhaps more important have been the advancements in ILT compute made possible by the GPU revolution. If you recall, we need to backwards calculate the mask design literally pixel by pixel. Modern chips have something like 80 billion transistors. A mask might have 2.5 trillion polygons and over 10 trillion corners of polygons, so it can take a very long time to run ILT and produce the backwards mask image, in some cases up to 30 million hours of CPU time for a full chip design. A data center with even tens of thousands of CPUs can take up to 10 days to do all that computation. Manufacturers expect this to be done in a day. So for years, ILT was reserved for dodgy hotspots in the design, where things are expected to be challenging to print. And they did not take it to the fullest extent. GPUs with their huge parallel processing capabilities have helped change that. In 2019, a software company called D2S debuted a software hardware solution that worked like a GPU CPU pair for full chip ILT. Then in 2023, Nvidia introduced a new GPU library of parallel algorithms called Culitho, accelerating certain workflows by 45 to 60 times. Nvidia remarked that with the library, 500 Nvidia H1 hundreds can do the work of four 40,000 CPUs and that certain mass designs that once took up to two weeks are now being done overnight. In early 2024, Nvidia, Synopsys and TSMC announced that they were moving Culitho into production. And now with the N2 node, they're rolling ILT and the Curvy masks into production for the first time with a few layers. There's perhaps one final thing that helped make N2's curvy masks possible. The chips themselves. From a 5G perspective, chips for AI accelerators are not all that technically different from chips made for other applications like mobile. But one major difference is in the quality of the work. The mobile phone market is very mature, and mobile phone SoCs are one of the priciest components inside the device. So mobile chip customers are very price conscious. AI chips like Nvidia's are different. The chips must be done to a very high standard, but there are enough margins to pay for it. Such a dynamic is justifying investment into advanced technologies like Curvilinear. Another such beneficiary is Photonics, which has received a shot in the ARM thanks to the problems of AI server scaling. These advancements, paid for by the boom, will filter down to other industries, benefiting them all. The crazy thing is that there's still so much more improvement ahead. TSMC's N2 masks are curvy. Yes. Yet they still haven't gone all the way. You can still sort of recognize the patterns. So ILT has not yet been fully leveraged to its greatest extent, but it's gonna get there and the results will be impressive. Alright, everyone, that's it for tonight. Thanks for watching. Subscribe to the channel. Sign up for the Patreon and I'll see you guys next time.