The topic of today’s discussion is: Why can’t the 28nm lithography machine achieve the 14nm or even 7nm exposure process through the multi-exposure process? Before understanding this complex semiconductor knowledge, you must first learn the prerequisite knowledge points. Today’s knowledge points include: First, what is the so-called 28nm lithography machine? How to define? Second, what factors determine the minimum accuracy of a lithography machine? What is overlay accuracy? Third, the actual parameters and definitions of transistors; fourth, detailed illustrations of the multi-exposure process/SAQP four-exposure process.
The reason for word count is divided into two paragraphs. Today I will talk about the first and second points, and the third and fourth points will be discussed next time.
The full text is more than 5,100 words.
1. How is the 28nm lithography machine defined?
First of all, there is no such name as 28nm lithography machine in the industry. This statement only exists among the people, but since it is popular science, it must be explained from the perspective of ordinary people.
So how should the so-called 28 lithography machine be defined?
According to a general understanding, if the minimum accuracy can meet the line width of 28nm, it can obviously be regarded as a 28nm lithography machine.
Then the question comes again, what is the actual line width of 28nm?
The internal structure of the chip is actually similar to a high-rise building. The spacing between the bottom transistors is the smallest, and the spacing between the metal interconnect layers above is relatively large. The so-called 28nm lithography machine should refer to the lithography machine that meets the characteristic size of the small transistors at the bottom of M0/M1.
What is the actual feature size of 28nm?
28nm has two classic versions of the process, namely 28HKMG (high-K metal gate) and 28Poly (polysilicon oxynitride gate). In fact, it is not just these two versions. TSMC had on the 28nm platform that year. The extension of LP, HPM, HPC, HPC+ and other versions is mainly to meet different customer needs. Some customers pursue frequency and performance, while others pursue lower power consumption ratio. Therefore, although they are all 28nm processes, there are slight differences in each process version.
The main reason why there are so many process versions is that changes in these processes will bring about changes in design rules.
On the one hand, more flexible design rules can reduce and improve photolithography steps. Second, different processes can significantly improve the gate spacing, with the purpose of improving performance or leakage.
According to the introduction of industry leaders, 28HPC/HPC+ and 28LP/HP/HPL/HPM are slightly different. The gate lengths are 40nm/35nm/30nm for the HPC version and 38nm/35nm/31nm for the LP version.
From this point, everyone discovered that the actual gate length of 28nm process is not 28nm. The largest is 40nm, and the smallest is 31nm. There are actually 6 specifications in total, and even 22nm can be considered a variant of 28nm.
Starting from the 40-28nm process, the nominal and actual gate lengths of the process nodes no longer correspond to one-to-one, but are relatively equivalent.
Therefore, in order to satisfy the 28nm process, without considering other process factors, the ultimate resolution of the lithography machine should be at least able to meet the minimum exposure line width of 40nm. However, since 40nm is not the main mass production process, the actual situation should be that the main process lithography machine should at least meet CD=35nm, which is what ordinary people usually understand as a 28nm lithography machine.
I checked the parameters of each model of ASML lithography machine. Theoretically, you can try the accuracy of NXT1950. After consulting Brother He, I did get a positive answer. Foreign FAB R&D engineers first tried to use 1950 for 28nm process development.
However, due to the many problems of 1950, it was quickly abandoned. And with the evolution of the process, the 28 mass-produced main lithography machines of foreign FAB became NXT1960B and NXT1970C.
Among them, 1970C is the most useful, but the domestic situation is different. At that time, the domestic 28nm progress was about 4-5 years behind the foreign advanced level. By the time it was really about to go into volume, 1970 was gone, so the domestic 28nm mass production version was actually NXT 1980D. Here is only an explanation of the actual situation.
In fact, doing the 28nm process in 1980 is equivalent to using a cannon to swat mosquitoes, because 1980 is much more expensive than 1970, but there is no way, there is no old cannon, so we can only use new cannons.
So based on the actual situation, we can make a conclusion here: The so-called 28nm lithography machine should actually refer to the NXT1970Ci model.
Lastly, two short stories are inserted. First, a certain big boss said that the HKMG process was caused by Intel that brought the world into a pit and took many wrong paths. The reason is that HK is said to be aluminum.
Second, as you can see from the above details, even if it is the same company, the same process platform, different versions of the process, the design rules are different.
In other words, If a chip design company wants to replace the original tape-out process today, it means almost completely overturning the back-end design and simulation work. The cost is very high. The more high-end the process, the higher the cost.
It doesn’t matter if the big IC design companies have money. Small companies really need to think carefully about whether they should change factories or whether they should change processes. This issue cannot be changed just by talking.
So don’t think it’s a simple matter for a design company to switch to a FAB factory for tape-out. It’s not like you go downstairs to buy groceries. If this one is not good, just change it. It’s complicated. Nothing is simple in the semiconductor field. Don’t use everyday thinking to understand things in semiconductors. In fact, it is not a concept at all.
When I have time, I will post a complaint about which expert on the Internet who knows nothing about XX is shouting every day that the expansion of TSMC’s Nanjing plant will take away the domestic FAB business. Damn it, do you think that switching to another FAB tape-out is just a matter of change? Are you going to grab the business as soon as you say it?
Second, what are the factors that determine the minimum accuracy of a lithography machine? What is overlay accuracy?
The lithography machine is a very large optical system. Any small internal or external errors will affect the final effect. In the world of light, if something is wrong, it is wrong, and if there is an error, there is an error.
If we look at it from the Rayleigh criterion formula: CD=K1*λ/NA. Obviously there are too many factors that affect the minimum CD. If we put aside the external influencing factors K1, K1 represents the polymerization degree, molecular weight, particle size, photosensitive agent of the photoresist, as well as the flatness of the silicon wafer, the incident angle of the light, and the amount of impurities/dust.
If the same platform and light source wavelength are compared together, then the main factors affecting the accuracy are the online measurement accuracy and the movement accuracy of the duplex stage.
Double workpiece stage, which is ASML’s unique “TWINSCAN” platform technology, is the biggest secret for ASML to remain competitive!
Insert a short story about ASML and Nikon. I once called it the three nails in Nikon’s coffin, and this is the second one.
was also mentioned in the book "Lithography Giant". Although ASML's PAS5500 also passed the certification of IBM/Intel on the 8-inch process, Intel rarely looked at ASML in actual procurement. The actual large-scale purchase on the production line was Nikon's S-204/205.
Purchase PAS5500 lithography machines in large quantities from OEM and storage customers such as TSMC, Samsung, and Micron.
In fact, in terms of equipment production capacity, PAS5500 is stronger than S205. Why doesn’t Intel use it?
The reason is very simple, because for Intel, which monopolizes the CPU market, the market pie is big enough and it can make money from it. It doesn't matter if the production is faster or slower, so it has never required high production capacity. At one time, Intel's capacity utilization rate was only 60%, and it did not even start work at night. Slowly and leisurely did not panic at all. Therefore, this is why even though the PAS5500 has a production capacity advantage over Nikon's S205, Intel was not interested in ASML in the early stage.
But TSMC is different. Production capacity is the lifeline. 60% capacity utilization? Stop work and no production at night? Isn’t this a waste of money?
For foundry companies such as TSMC, they must have advantages in cost, efficiency, and production capacity in order to break out of this fierce battlefield and gain a firm foothold in the competition.
So competition makes people progress, and monopoly makes people lazy. ASML uses high-capacity lithography machines to firmly capture the hearts of TSMC. The same goes for storage factories. Competition is very fierce. How can you make money lying down like Intel?
But having said that, FAB factory work is really cruel, and only engineers with tolerance and absolute obedience cultivated in the East Asian cultural circle can adapt.
In the past, TSMC passed on a joke: You will be very rich working at TSMC, because you have no time to spend money at all.
The cruelty is evident, even 996 is weak in front of FAB!
FAB’s work and system are simply not done by humans! Recently, there is a franchise buddy on the Internet. In a long series, his story about the working manager of the franchise in Singapore is called "Review of a Chip Veteran". You can check it out. It is very realistic.
Perhaps after reading it, you will know why Sajia dared to make the bold statement "If TSMC's new Arizona factory does well, I will go to the Colorado River to wash my hair upside down".
Because I think it’s impossible for people in the United States to do this job well. Why don't you go to Wanwan and recruit people to come over? But everyone knows that those who want to go to the United States are just to get a green card. Those who have ulterior motives are really evil if they do well.
Back to the topic, how does ASML increase production capacity? What kind of unique skills does it have?
Looking back, the solution was actually quite simple.
Before the pattern is exposed to the wafer, the wafer must be accurately measured. Both measurement and exposure take time. In order to reduce the time required for each process and improve efficiency, why not start measuring and aligning the next wafer while exposing one wafer?
In this way, the TWINSCAN system was born
TWINSCAN system: One is responsible for pre-alignment measurement, and the other is responsible for exposure
TWINSCAN is the first and only lithography system with a dual-wafer work platform.
Wafers are alternately loaded onto the TWINSCAN platform. When the wafer on one platform is being exposed, the other wafer is loaded onto the No. 2 platform for alignment and measurement. Then the two platforms exchange positions. The wafer originally on the No. 2 platform is exposed, and the wafer on the No. 1 platform is unloaded. New wafers are then loaded, aligned and measured.
This parallel solution of measuring, aligning and exposing at the same time can greatly increase the production capacity of the lithography machine per hour, which helps TSMC greatly improve production efficiency and enhance final benefits.
In 2001, the first TWINSCAN dual-wafer platform system using this revolutionary technology was shipped - TWINSCANAT:750T lithography machine.
750T lithography machine uses a KrF light source system with a wavelength of 248nm and supports the production of 130nm process.
Soon, ASML’s i-line lithography machine also introduced a dual-wafer platform, namely TWINSCANAT:400T; then this technology was introduced to a higher-end 193nm ArF lithography machine, namely TWINSCANAT:1100. Therefore, from the i-line to the KrF line, the TWINSCAN system spans lithography machines of various ASML platform models, expanding the scope of technology and allowing all chip layers to be exposed on the new platform.
ASML's continuous innovation capabilities provide incremental improvements to the resolution, overlay accuracy and productivity of the TWINSCAN platform - in different ways such as platform upgrades, new system upgrades and on-site upgrades, in a word, whatever makes customers comfortable.
Therefore, the movement accuracy of the double workpiece stage is, to a certain extent, the key to the alignment accuracy of the lithography machine!
Understanding the working process of the double workpiece table is incredible. This is the power of technology.
They are moving at high speed all the time. When stationary, they accelerate rapidly and then stop suddenly to reach the position where they should stop. The accuracy is breathtaking.
If calculated based on the instantaneous acceleration, it has exceeded the launch speed of the rocket. It will stop at the exact position at the next moment without any mistakes, because any mistakes at this speed cannot be compensated.
If you make a mistake, although the entire wafer will not be scrapped and you have to start over, but if you make this mistake many times, run away quickly, because the engineer will come to chop people with a 40-meter machete.
The double workpiece table accelerates - emergency stop - accelerates - emergency stop, repeating this process while maintaining long-term stable working status.
So in a sense, the dual workpiece stage system determines the maximum production capacity of the lithography machine, as well as its accuracy. In the world of lithography machines, this is called Overlay - overlay accuracy.
Taking NXT1980Di as an example, the official parameters are OPO≤3.5nm, DCO≤1.6nm, MMO≤2.5nm. Usually worse 4-5
Here’s the point! Popularize knowledge that 99% of people don’t know.
OPO means OnProductOverlay. The overlay accuracy on the product, because the chip manufacturing process is somewhat similar to the process of building a building, is equivalent to the alignment accuracy between the last exposure and the current one. This accuracy is within 3nm.
DCO is the abbreviation of DedicateChuckOverlay, which is equivalent to the same device having its own accuracy. This is within 1.6nm.
MMO is the abbreviation of Mix-and-MatchOverlay, which is equivalent to the overlay accuracy between different devices. This can be less than 2.5nm.
Do you remember the multi-exposure process? A very important step in the multi-exposure process is to split the original mask pattern into two pieces and expose it twice to obtain a smaller pattern.
Obviously, whether it is OPO, DCO, or MMO, these parameters together determine whether you can use the multi-exposure process, as well as the consistency of the exposure pattern after mass production, and the final alignment accuracy of the upper and lower layers.
If any parameter is not enough, then the graphics produced by multiple exposures must be crooked, horrible, and a mess. Not to mention the yield, the entire wafer will probably be scrapped.
This picture was posted on the Internet today:
The picture is provided by a group of friends. If there is any infringement, let me knowTAGP H26
I can only say that the above is full of errors. The NXE3400 is written as NXT, the NXT2000 has been discontinued long ago and has never been sold in China, and the 2050 can achieve 5nm. There are too many mistakes.
I will go directly to the data compiled by Teacher Guan:
Let’s find faults and see how many places are wrong.
Back to the topic.
The reason why 1970 can only do 28nm, 14nm is very difficult, let alone 7nm, the reason is here, compared with 1980, the Overlay of 1970 is far behind!
Overlay’s performance is not good enough, I can’t do it!
So the double workpiece stage, to a certain extent, is the most important key component besides the light source and objective lens system.
Its stability, alignment accuracy, and mean trouble-free running time directly affect the actual working status of the lithography machine, and even affect the process level and production capacity of the entire FAB.
Therefore, up to hundreds of units (Field) need to be exposed on a wafer, and advanced photolithography machines can expose more than 300 silicon wafers an hour, while ensuring that the exposure amount is the same every time.
Assuming that there are 300 unit areas that need to be exposed on a 12-inch wafer, it is equivalent to 2.16 million exposures a day and 788.4 million exposures a year. The stability and effect consistency of the dual workpiece stage and the entire equipment is a huge test.
Maybe these numbers don’t make you think anything, but after thinking about it, the technical content represented by these numbers is really shocking.
Someone once compared it to two planes flying at high speed. One of them took out a knife and carved words on the other plane on an area the size of a grain of rice.
It is the most difficult challenge in engineering for a machine with such outrageously precise movements to maintain stable operation 24/7, with countless technical peaks to overcome.
Previously, it was claimed that a certain laboratory in China could achieve several nanometers, and a lot of people touted that surpassing ASML was just around the corner. You must know that there is a huge difference between laboratory equipment carving two straight lines and commercial equipment running stably around the clock to expose complex graphics.
I know that China has made breakthroughs in key technologies in some aspects, but there is still a long way to go and there is still a long way to go before real success.
Visit:
Jingdong Mall