1000 days for Huawei chips

Written by: Rick

In the last days of August 2023, Liu Ye clearly felt that the atmosphere of the chip department had become a little different.

Colleagues began to shout that they were going to return strongly this time, and the optimistic atmosphere spread throughout Huawei.

The boiling moment of all staff entered a high point at noon on the 29th, and Liu Ye found that the circle of friends was rarely brushed by colleagues - Huawei launched the new flagship mobile phone Mate 60 pro in its official store without warning.

And what people are most concerned about is: How did the Kirin 9000s equipped with the **Mate 60 pro come about? **

It will soon be the birthday of the Kirin 9000, and before that, we want to tell you the story of its next generation chip Kirin 9000S, which is an inspiring story.

To this end, the editorial department found people related to the semiconductor industry, former employees of Huawei, and people from upstream and downstream enterprises in the supply chain, hoping to get some answers, and thank them for sharing as frankly as possible in a limited scope, so that we can get as accurate a glimpse of the development path of the chip industry in the past three years as possible.

Among them, Zhiwei felt that not only Huawei, but the entire industry chain was working together for a common goal to complete this result.

Codename, Charlotte

In fact, since the launch of the Kirin project, many chips have been staggered and developed.

For example, the Kirin 920 and Kirin 910 were developed and delivered almost in parallel, and this method is internally known as the “towel twisting mode”. **

Then, as is customary, when the Kirin 9000 is in mass production in 2020, there should already be parallel new projects in development.

Wu Xu, a relevant person in the chip industry contacted by Zhiwei, disassembled the Huawei Mate60 Pro purchased on August 30 and decaped the chip.

In addition to observing and analyzing the internal structure of the chip, Decap is also carried out to find the real mass production date of the Kirin 9000s. Previously, it was rumored on the Internet that the “2035CN” on the chip shell represented that the chip was produced in the 35th week of 2020, but he believed that this information had little reference value, ** more like some kind of confusing “camouflage”. **

Wu Xu obtained a special code “2017” after pickling and amplification, and after several verifications, he believed that **This is the TO (tape-out) date, that is, the date of the final step of integrated circuit (IC) or printed circuit board (PCB) design **, generally speaking, this number appears in the 13~15th layer of the chip metal layer.

**And “2017” means, week 17 of 2020. **

Generally, the chip will begin mass production 100~200 days after finalization, so the person believes that the real production date of the Kirin 9000S in his hand is: early 2021. **

Before mass production, chips go through four stages, design stage, development stage, trial production stage and mass production stage. Tape-out is often the most important and money-burning link in chip manufacturing, and some chip factories have estimated that a tape-out of the 7nm process costs $30 million. This process lasts at least three months (including raw material preparation, lithography, doping, electroplating, packaging and testing), and goes through more than 1,000 processes, with a long production cycle.

**Combined with the tape-out time and the production date of the chip, the person judged that the Kirin 9000s was established at least no later than 2020, and the purpose was not to be produced at TSMC from the beginning. **

Another Huawei employee confirmed to Zhiwei that the production time of the Kirin 9000s was about one and a half years, ** time “about the end of 19 years”, and some energy was spent in the design stage.

The person said that HiSilicon is different from other chip design factories in that “basically they will do DTCO (circuit design and process co-optimization), and the details are decentralized to the transistor level, not just simple wiring.” This has the advantage of better performance, with the disadvantage of taking longer and higher technical design.

For example, the density and performance of the normal design is 95%, and after DTCO optimization, it may reach 100% or better, but it is time-consuming and requires design collaboration with FAB vendors. Various chip design factories can do it, but basically do not do it, and Qualcomm sometimes does some. ”

According to the information it has, there was a version inside the Kirin 9000s called Hi36B0. **Hi stands for Huawei HiSilicon, 36 represents Kirin flagship product line, and B0 stands for 11th generation. In the mass production of this chip, a new logo is adopted, that is, Hi36a0V120 instead of “b0”. The 2 and 0 in the following “V120” represent version changes and small optimization iteration numbers (the 1 after V refers to the product generation on other Huawei chips, for example, the first generation of the chip of the TV is V100, and the second generation is V200, but the meaning is not sure on the Hi36 Kirin series).

**In addition to this string of code names, the Kirin 9000s also has a more memorable name inside, Charlotte, the name of the American city. **

Although the Kirin chip is known as a Chinese mythical beast, the specific model has always been named after an American city internally. The previous generation Kirin 9000 was Baltimore, 990 for Phoenix, 985 for Tucson, 980 for Atlanta, and 970 for Boston.

In terms of process, from the SEM (scanning electron microscope) diagram of the Kirin 9000s obtained by the Zhiwei editorial department, the Cell Height of the Kirin 9000s (standard unit height, commonly used to measure the chip process level) is 240nm.

After pickling, the local map of Kirin 9000s was magnified 600,000 times

In 2020, when TSMC disclosed its original 7nm low-power and high-density solution, the Cell Height value was also 240nm.

That said, there is no doubt that Huawei’s Kirin 9000S has reached the 7nm process level. **

A neatly arranged Kirin 9000S transistor after 100,000 times magnification

At the same time, the editorial department obtained the chip physical structure diagram of the Kirin 9000S, and the structure of the Kirin 9000S is very different from the previous generation chip Kirin 9000.

Therefore, we can tell you with a little excitement or pride here: **Kirin 9000S is a brand new chip that is not modified from the Kirin 9000 inventory and has reached 7nm advanced process. **

Wu Xu told Zhiwei that Charlotte has a total of 8 cores, which are three clusters (a way of arrangement), the distribution is 1+3+4, the main frequency is up to more than 2600MHz, and the GPU is Maleoon 910.

Huawei’s 5G baseband part has always been a design with 4G + 5G two modules connected with a Barong baseband chip in the middle, and this generation does not use this bridging method, but integrates 4G and 5G with one module.

Compared with the Kirin 9000, the huge area of Charlotte CPU Cluster has undergone big changes, and the bus of this generation, unlike the previous generation of buses and super large cores, uses performance libraries, and this generation only uses performance libraries for super large cores.

In terms of GPU, Charlotte’s Maliang 910 is a Cu design. Its design scale is slightly smaller than the previous generation, for about 4CU two groups of ALU Core, each group of 128Alus, a total of 2x4x128Alus=1024Alus, the frequency is up to 750Mhz, the theoretical performance is 1536Gflops, and the middle is GPU L2 Cache, which is about 1MiByte. **In terms of GPU specifications, it is not the same as the common IMG/MALI/Adreno/Rdna/Cuda. **

However, as we all know, Huawei does not have the manufacturing capabilities of advanced process chips, so the question arises:

In the case of multiple rounds of sanctions, how did Huawei, or Chinese manufacturers, make 7nm chips? **

White Knight

In the past, Huawei trusted TSMC, ** a relevant person revealed to Zhiwei that at that time, Huawei’s senior management had judged that the possibility of TSMC cutting off supply was low. **

On the one hand, before the sanctions, the two sides have reached a cooperation in the production of the most advanced process 5nm process Kirin 9000 chips, and are in a situation of continuous in-depth cooperation. On the other hand, chip foundry anchoring a factory manufacturing is also due to cost considerations.

“Now it seems that in that environment, it seems unwise to insist on putting eggs in one basket, once TSMC rejects Huawei’s tape-out (trial production), it will not be able to continue production, and go through the following process,” the relevant person said to Zhiwei.

In May 2020, the United States escalated sanctions, which announced restrictions on manufacturers using American technology (such as TSMC) to Huawei foundry chips, this ban was not immediately implemented, and the United States gave a 120-day buffer period.

At the results conference on July 16, 2020, TSMC chose to compromise, indicating that after September 14, TSMC will no longer continue to supply chips to Huawei.

Huawei’s response was very rapid, and immediately after the sanctions were issued, the decision to mass-produce the Kirin 9000 was immediately issued internally.

Generally speaking, HiSilicon designed chips to go through multiple wafers (after design, they are sent to the factory for trial production testing), and an employee of Huawei’s relevant department told Zhiwei that **At that time, the decision of the Kirin 9000 was originally to be cast 3 times, but it encountered a sanction order after the second time, so “the third time was not invested, and it was directly mass-produced”. ** These chips helped Huawei support nearly two years after a complete cut.

On October 31, 2020, the Kirin Chip and Technology Development Department held a Kirin 9000 press conference, the core theme of which is ** “Firm Faith, Never Give Up”. **

Photo courtesy of interviewee

However, the Kirin 9000 is one less piece with one, and who will build the future chip Kirin 9000S?

2020 is a special node, Chinese manufacturers are in a moment of survival not only Huawei, but also SMIC.

This year’s Mid-Autumn Festival, which coincided with the National Day holiday, former SMIC employee Xu Qin and his team colleagues were suddenly called to the company, and they received the shocking news that the Bureau of Industry and Security (BIS) of the US Department of Commerce had sent letters to some of SMIC’s suppliers in accordance with US import and export control regulations, requiring them to apply for export licenses before supplying to SMIC.

The news was only officially released by the U.S. Department of Defense on December 4, which announced that four Chinese companies, including SMIC, were officially added to the list of “military end-users.”

At that time, SMIC was the Chinese mainland most likely chip manufacturer to become one of the world’s most popular manufacturers, and if it could not obtain advanced equipment and raw materials from abroad, SMIC’s growth progress would be seriously slowed down.

The panic of surprise and the work of emergency adjustment go hand in hand. **“It asks everyone to analyze their current device, what if it stops?” What is the solution? Parts, raw materials, equipment that needs foreign manufacturers to come to do service, can you do it yourself, how much can you do? ** Xu Qin recalled.

" The worst plan is to cut completely and not communicate with each other. ”

Correspondingly, relevant American companies are also interpreting the information released by the US government with the legal team, but the laws and regulations involving national interests can only be implemented through friendly negotiations and immediately, and cannot cross the thunder pool by half a step. After a short period of panic, SMIC found that the restrictions focused on the technology and equipment required for high-end processes, and the “stuck neck” left a breath. Therefore, the pace of domestic substitution has been pushed into the fast lane.

However, the most affected is SMIC’s advanced process team, according to people close to SMIC, **There has been an internal proposal to purchase ASML’s EUV lithography equipment ** (commonly used in 7nm and below process equipment) first, while developing related process technologies.

**However, this proposal was not accepted. **Because at that time, both TSMC and Samsung first used DUV lithography to complete the “transitional version” of the 7nm process, and only after accumulating more experience and reaching a certain scale did EUV be introduced. ( DUV lithography machine accuracy is lower than EUV equipment, it is generally believed that the “5nm” process is its manufacturing limit, but the industry will use EUV lithography machine around 7nm)

Another part of the reason is that the equipment is too expensive, the order time is delayed, and subsequent deliveries are constantly jammed, and it has not been delivered.

SMIC originally planned to march from 28nm to 20nm, but later internally decided to abandon 20nm and go directly to the more advanced 14nm. In 2019, the pilot production yield rate was rapidly increased from 3% to 95%, achieving mass production.

As for the development phase of the 7nm chip, we can see it in the December 2020 letter from Mengsong Liang (currently SMIC’s co-CEO) to the board of directors. “During this period (2017~2020), I worked hard to complete the development of technology from 28nm to 7nm, a total of five generations… At present, 28nm, 14nm, 12nm, and n+1 technologies have entered mass production, and the development of 7nm technology has also been completed, and it can immediately enter risky mass production in April next year (2021)…”

Interestingly, the estimated risk mass production time in the letter is April 2021, which is surprisingly consistent with the production time of the Kirin 9000s judged above. **

The new question is, which technology does SMIC use in the absence of advanced lithography machines? If the mass production of 7nm process is on domestic chips, what problems will be encountered?

Draw thin lines with a brush

We need to re-understand the chip, and the thin chip may actually have as many as a hundred layers inside.

The chip process is to first make the transistor morphology on the silicon chip, layer by layer deposited coating, piled up the upper metal layer, isolation layer, passivation layer, of which the bottom is the most core, the most cutting-edge part of the process, the capacitor and transistor are here, called the underlying device. Generally, the nanometer chip we refer to refers to the bottom transistor part.

Below 28 nanometers, because of the serious quantum tunneling effect, there will be leakage, and the planar transistor cannot meet the requirements of use, and the gate must be wrapped up like a fish fin to make FinFET, that is, “fin field effect transistor”. Speaking of which, this innovation came from Professor Hu Zhengming, a Chinese scientist and former chief technology officer of TSMC.

At this time, the stereoscopic transistor is actually difficult to quantify with length, to see what process level it reaches, that is, commonly known as a few nanometers, depending on multiple technical indicators, such as transistor gates, minimum spacing between fins (Fin Pitch), Cell Height, and transistor density (how many transistors can be accommodated per millimeter on the chip).

The state-of-the-art 193nm DUV immersion lithography machine can provide a half-cycle resolution of 36~40nm, meeting the requirements of 28nm logic technology nodes. Smaller than this size, double or even multiple lithography is required.

The core of multiple lithography technology is to split the original layer of lithography into two or more masks, and use multiple lithography and etching to realize the original layer of design of the pattern, so that it can etch more than a single exposure CD (Critical Dimension, refers to the design of a special line pattern reflecting the characteristic line width of the integrated circuit in order to evaluate and control the graphic processing accuracy of the process in the integrated circuit photomask manufacturing and lithography process).

Double exposure is widely used at 22nm, 20nm, 16nm, and 14nm technology nodes as well as advanced process non-critical layer fabrication. But after the EUV lithography machine technology matured, TSMC and Samsung gradually used EUV lithography machine, which is a completely different technical route, only one exposure can achieve the effect.

SMIC wants to achieve 7nm without EUV lithography machine, which can be said to use “old technology old machine” to achieve advanced goals, which is a bit like carving with an iron pestle. **In 2019, TSMC produced 7nm node (N7) chips through DUV equipment, and then began using EUV lithography machines.

There are many technical routes to achieve double or even multiple lithography, such as LFLE process, LELE process, LELELE process, SADP, SAQP technology, etc.

It has previously been reported that Huawei may use the so-called “chip stacking” technology to achieve the effect of 7nm chips with two 14nm chips. But a person who understands the chip process told Zhiwei that this is unlikely, " Generally, this process is used for HBM (high bandwidth memory) 3D packaging technology, not a 14+14=7 problem, to solve the routing design, energy consumption, area and other problems between the two chipsets are hugely difficult, and it is completely unrealistic to use in mobile phone chips. ”

A relevant person told Zhiwei that SMIC adopted the SAQP technology route to realize the 7 nm process.

Another person close to SMIC revealed that when Liang Mengsong joined SMIC in 2017, he required all the technicians in the department he was responsible for to learn SAQP technology, “New engineers must work overtime to learn this technology.” ”**

So what is SAQP technology?

The Chinese name of SAQP is “self-aligned quadruple exposure”, and its implementation principle is simple and popular:

(1) First draw the “lattice” with a lithography machine, and then use an etching machine to engrave the “lattice”;

(2) Chemical vapor deposition coating on the engraved lattice;

(3) The coating on the horizontal plane is removed by etching technology, at which point we obtain a “sidewall” composed of a thin film;

(4) another round of etching, so that we obtain a denser “lattice” composed of the side walls of the film;

(5) Chemical vapor deposition coating again;

(6) Use etching technology to remove the coating on the horizontal plane;

(7) Etch again to obtain a more encrypted “lattice”;

(8) Under the block of the lattice, continue to etch downward;

(9) Remove the coated lattice, leaving the “lattice” that is really needed.

Zhiwei also rendered a GIF for everyone to better understand:

So far, we have used etching technology to draw thin lines with only a very thick “brush” such as a DUV lithography machine.

In fact, no matter which technical theory mentioned above, it has been around for many years, but in the technical selection and node selection of the process, the learning curve will be extremely important, because each step requires a lot of money and manpower.

And SMIC’s ability to accomplish such a difficult thing, in addition to key technical personnel, may be related to its corporate culture.

Xu Qin believes that “obedience, strong execution, and absolute pragmatism at the technical level” have created SMIC, which has a history of more than 20 years.

**" After clarifying the R&D goals, it is result-oriented, 100% implemented, and more respectful of the people who do things. ** According to his observations, the change of personnel has little impact on the research and development of the company’s various projects, coupled with strong execution, so that the company can have a good development.

** Zhiwei learned of an unconfirmed industry rumor that SMIC’s advanced process team had been on a year-round day off for three consecutive years, with only one day off on New Year’s Day. **

From the results point of view, according to the time node of the past advanced technology, **SMIC did take 3 years to complete the road of other manufacturers for 10 years. **

Combination, yield, success or failure in one fell swoop

Relevant people revealed to Zhiwei that after Charlotte established the project, the original foundry was set as SMIC, and this was the only feasible solution, when Huawei was in the stage of being surrounded by technology, and the advanced chips purchased from TSMC were about to be exhausted, and material imports were also hindered.

It is worth mentioning that during the development of the Kirin 9000 chip, Huawei had a chip in SMIC, “but then it was not on, but the next generation (9000s) is,” a SMIC employee mentioned. **

Under the pressure of sanctions, de-A (Americanization) is in full swing within Huawei. **" Not only technical A, office software, professional software is the same, do not do it yourself, and finally reach the US products and technologies completely exit the workflow ** A former employee mentioned that at that time, Huawei’s communications department directly dismounted to re-demonstrate the feasibility.

Since it is impossible to judge the extent of the tightening restrictions, it is a top priority to complete the mass production of Charlotte in the shortest possible time. The first step in cooperation is process migration and matching, which is often overlooked.

Generally speaking, in the advanced process, the design scheme and each foundry there is also a process of adaptation, TSMC, Samsung and other advanced process foundries have a special team for “transcoding adaptation”, but " SMIC did not have such a design rule migration team at that time, Huawei had sent a team to adapt the process " Relevant people said that the whole process took about 3~6 months.

After that, yield becomes key.

In the semiconductor field, yield is related to the cost of mass production of chips, and the more qualified quality chips on each wafer, the lower the cost of the chip. The final yield is composed of the product of each process, even if we assume that each process on a fab line is as high as 99%, then after 500 processes, the overall yield is only 0.66%, and the result is completely waste. In general, yield can be subdivided into Wafer (silicon wafer) yield, Die yield, and package yield, and Die yield has a greater impact on the total yield.

Relevant people told Zhiwei that Charlotte’s yield is about 35% when it is risk mass production, and in general, it is necessary to achieve mass production at least 50%, but this is also double the process cost that can reach more than 90%.

Zhiwei also learned that this year, a packaging factory received an order for Charlotte chips,** the factory has reached a monthly production capacity of 4 million pieces in recent months. **

As for the true total yield now, we do not know, because it is strongly related to the cost of the chip, which is generally regarded as a secret by manufacturers.

However, relevant personnel revealed to Zhiwei that **Charlotte’s yield has reached about 50%-60% in the early stage of formal mass production, and the yield climb after that is also considerable, which can support its large-scale production with controllable costs. **

So, you can see the news: Huawei aims to ship 60-70 million smartphones in 2024.

In 2022, Huawei’s smartphone shipments will only be about 30 million.

At this point, perhaps we can breathe a long sigh of relief and say:

** The light boat has crossed the Ten Thousand Heavy Mountains. **

Postscript

The success of the Kirin 9000s may be a milestone in the localization of chips, but this is only a phased victory on the long road. **

A semiconductor industry practitioner worriedly said to Zhiwei that after the results were shown, it is expected that the future sanctions will be more violent, and this success is to “breathe” in the limited space under the sanctions, **" stuck in the neck this thing, this time it is stuck here, next time? Next time you may reach deeper." **

When doing this article, Zhiwei really felt that technological innovation breakthroughs are more the result of collaborative operations, and when a crisis leverages the industry, it is impossible to simply judge whether this is a catastrophe or a chance to die. Many practitioners have an inexplicable “belief”. In their eyes, as long as the goal is determined and pragmatic and coordinated, there is nothing that cannot be accomplished.

We thought, this is roughly what it is called:

**Faith can move mountains. **

The next generation of chips, codenamed “Nashville,” is on its way. **

( At the request of interviewees, the names of the persons mentioned in the article are pseudonyms )