Semi.org has made their predictions for silicon wafer shipments (area) through 2025, with growth of 4.8% this year and growth each year through 2025 except 2023.  Simply put, macro conditions in 2023 are blamed for the decline in polished wafers, but growth resumes again in 2024 and 2025, which we have to assume means that inflation is brought under control in 2024 and beyond.  Hopefully, that will be true, but we expect estimates to come down a bit further  for 2024 & 2025 as we move through 1Q 2023.

​Taiwan based Foxconn (2354.TT) is a massive employer in China with 12 factory complexes, the largest of which is located in Longhua Town in the city of Shenzhen, which has come to be named “Foxconn City” and covers 1.2 mi2 within its walls.  Many workers live in the dormitories and walk to the 15 factories within the ‘city’, which has its own stores, restaurants, banks, and hospital, and while numbers vary considerably, the lowest worker count we have seen is ~200,000.  Foxconn also has a similar ‘city’ in Zhengzhou, said to also employ ~200,000 workers, which is also known as “iPhone City”, as much of Apple’s (AAPL) iPhone line is assembled at the Zhengzhou complex.
Late last week the local government informed the management of the Zhengzhou complex that the city would be placed on lockdown due to an outbreak of COVID-19, limiting travel.  This forces workers to stay ‘on campus’ in the dormitories, as the factories are allowed to continue production under a ‘closed loop’ system, meaning no one leaves the complex.  As this was a situation that developed rapidly, it seems Foxconn was caught off guard and did not have the resources needed to feed and house the entire staff of workers.  Under lockdown rules workers were not allowed to eat communally in the company cafeteria, and were forced to eat in the dorms, which caused conditions to be difficult early on. 
There have been a number of news stories and videos about workers jumping over fences in order to leave the plants for fear of being caught in a COVID outbreak or unhappy about the poor living conditions, so Foxconn began to increase weekly bonuses to workers from ~$14/wk. to ~$55/wk., and for those that worked for more than 25 days, the bonus was increased from ~$205 to ~$685.  Those that put in “full effort”, meaning remained on campus for the entire period, could earn a potential ~$2,054 bonus, against a typical worker salary of between $410 and $550/month.
Over the last two days there have been videos circulating social media saying that there have been 8 deaths at the Foxconn Zhangzhou plant during the lockdown, which Foxconn denies under the premise that the videos were ‘doctored’, but the real question is how much the issues at Foxconn will affect iPhone 14 family production.  Recent estimates indicate that the plant’s output could be reduced by as much as 30%, due to staff shortages caused by workers leaving, but some of that shortfall can be transferred to the Shenzhen facility, which faced two closings this year in March and July but is currently not under lockdown. 
Quick math based on historic Apple iPhone shipments would indicate that Apple would typically ship 85.97m iPhone units in the 4th calendar quarter, as each quarter this year has been up between 2.5% and 3.2% y/y.  Taking the mid-point of that spread generates the 85.97m units for 4Q (cal.).  As a check, 4Q shipments for the iPhone typically represent 36.5% of the full year’s shipments, and the number above would represent a slightly more conservative 35.6% of full year, so it is certainly within reason.  As Foxconn assembles ~70% of iPhones, that would represent 60.17m units in 4Q, or 20.06m units/month.  Assuming the lockdown lasts for one month and that Foxconn’s production rate is reduced by 30%, that would represent a shortfall of 14.04m units for the quarter, or 16.3% of 4Q (cal.) unit volume.  Again this assumes a one month lockdown, a 30% volume reduction, and no shift to another Foxconn plant, all of which would likely be the worst case scenario. 
We would expect that with the unusually large ‘incentive’ payments Foxconn is offering, the volume reduction will be less than 30%, and while the lockdown could last a month, we expect it will last 2 to 3 weeks, as the company and the city are testing virtually everybody on a regular basis.  We expect Foxconn can offload ~10% of the ‘missed’ volume to the Shenzhen plant, and, in worst case, Apple could shift another 10% to other assemblers, bringing the shortfall to ~11.2m units spread across the globe.  While this is still a substantial shortfall, iPhone buyers have faced order fulfillment delays before, with two weeks not being uncommon during periods of high demand.  As Apple has been the only smartphone brand to see shipment growth last quarter, iPhone 14 family customers could see longer than usual delivery delays, but good logistics could shorten that time a bit by pulling inventory from slower sales regions.  All in the worst case scenario for iPhone 14 shipments and delivery is likely less of an issue than the news stories seem to hint at, and even our quick calculations assume a positive y/y iPhone shipment number, which is certainly not a guarantee given the macro environment, so unless the lockdown is sustained or spreads to Shenzhen, we expect it will do little to hinder Apple sales for the 4th calendar quarter, and in worst case push some shipments into 1Q ‘23

​As we have noted over the past month or so, the Biden administration has been developing a set of new trade restrictions on semiconductors that are designed to extend those already imposed, with the intent of limiting China’s ability to compete in the world semiconductor markets and reduce China’s ability to develop a number of technologies, which the US government feels could be used to advance China’s military capabilities.  On Friday the US Bureau of Industry and Security/Office of Congressional and Public Affairs, a part of the US Department of Commerce, issued a statement on the implementation of new export controls on advanced computing and semiconductor manufacturing to the PRC and an additional 31 entities to the “Unverified List”, essentially a list of companies that the DOC says remain ‘unverified’ as to their ‘bona fides’ due to foreign government non-compliance.  Companies on the “Unverified List” get moved to the Entities List if they remain unverifiable for an extended period of time.  Nine entities on the “Unverified List” were removed, having met the requirements.
The new rules are as follows:

1. Adds certain advanced and high-performance computing chips and computer commodities to the Commerce Control List (CCL)
2. Adds new license requirements for items destined for a supercomputer or semiconductor development or production end use in the PRC – Supercomputer defined as 100 or greater 64-bit Petaflops od 200 or more 32-bit petaflops that is in a 4’ x 4’ x 6.5’ or smaller rack
3. Expands the scope of foreign-produced items subject to license requirements to twenty-eight existing entities on the Entity List that are located in the PRC (below).
4.  Adds certain semiconductor manufacturing equipment and related items to the CCL (below)
5. Adds new license requirements for items destined to a semiconductor fabrication “facility” in the PRC that fabricates ICs meeting specified items.
   1. Licenses for facilities owned by PRC entities will face a “presumption of denial,”
   2. Facilities owned by multinationals will be decided on a case-by-case basis. The relevant thresholds are as follows:

i.Logic chips with non-planar transistor architectures (I.e., FinFET or GAAFET) of 16nm or 14nm, or below
ii.DRAM memory chips of 18nm half-pitch or less

1. NAND flash memory chips with 128 layers or more.
2.  Restricts the ability of U.S. persons to support the development, or production, of ICs at certain PRC-located semiconductor fabrication “facilities” without a license – Support is defined as shipping, transmitting, or transferring (in country).
3. Adds new license requirements to export items to develop or produce semiconductor manufacturing equipment and related items
4. Establishes a Temporary General License (TGL) to minimize the short-term impact on the semiconductor supply chain by allowing specific, limited manufacturing activities related to items destined for use outside the PRC.

The new rules are a bit specific as to what equipment has been added but the idea is that any tools that aid in the development of semiconductors that are above ‘generic’ levels are being included, specifically those related to cobalt, which has promise in reducing the power requirements of semiconductors, an important part of supercomputers where massive numbers of processors are used.   While these rules began going into effect on October 7, with restrictions on personal development support beginning on 10/12 and advanced computing rules on 10/21, we expect the list of tools to continue to expand as the DOC delves further into potential advances in semiconductor process technology that can be used to advance supercomputing and AI.
While the exact language is not specified, exceptions to the ‘assumed denial’ rule that covers entities and equipment specified in this and previous orders, license exceptions (aka “Temporary General License”) will be given to “minimize the short-term impact on the semiconductor supply chain by allowing specific, limited manufacturing activities related to items destined for use outside of the PRC.”, which we assume to mean that non-Chinese companies operating in China will be able to produce and export semiconductors to customers outside of China, although the ‘temporary’ wording would suggest that the US government is encouraging those manufacturers to move production to the US or other locations outside of China.  Discussions between South Korean (Samsung Electronics (005930.KS) and SK Hynix (000660.KS) and the US government have been ongoing to clarify specific issues given their production locations on the Mainland.
The full rule (146 pages) is here:  https://public-inspection.federalregister.gov/2022-21658.pdf
Chinese semiconductor companies on the entities list (partial):

• Beijing Institute of Technology
• Beijing Sensetime Technology Development Co., Ltd. (0020.HK) (Case-by-case review)
• Changsha Jingjia Microelectronics Co., Ltd. (300474.CH)
• Chengdu Haiguang Integrated Circuit
• Chengdu Haiguang Microelectronics Technology
• China Aerospace Science and Technology Corporation (CASC) 9th Academy 772 Research Institute
• Dahua Technology (002236.CH)
• Harbin institute of technology
• Higon (pvt)
• IFLYTEK (002230.CH) (Case-by-case review)
• Intellifusion (pvt) (Case-by-case review)
• Megvii Technology (pvt) (Case-by-case review)
• National Supercomputer Center Zhengzhou
• National Supercomputing Center Changsha/Guangzhou, Jinan, Shenzhen, Tianjin, Wuxi
• National University of Defense Technology
• New H3C Semiconductor Technologies Co., Ltd (HPE JV)
• Northwestern Polytechnical University
• Shanghai High-Performance Integrated Circuit Design Center; Sugon
• Sunway Microelectronics
• Tianjin Phytium Information Technology
• Wuxi Jiangnan Institute of Computing Technology
• Yitu Technologies (pvt) (Case-by-case review)

 
List of Added Semiconductor Manufacturing Equipment (Partial)
 

• Equipment for depositing cobalt through electroplating processes
• CVD for cobalt or tungsten deposition (<3nm)
• Any tool that can fabricate Tungsten metal contacts between 100⁰C and 500⁰C
• Any tool that conducts a plasma process that includes Hydrogen, Oxygen, Cobalt, Tungsten, and Ammonia derivatives at certain temperatures
• PVD deposition tools for Cobalt at 10nm or less
• Atomic Layer Deposition tools depositing organometallic aluminum and Titanium Halide
• Any tool capable of depositing Titanium Nitride or Tungsten Halide
• Sputtering tools for Cobalt deposition
• Tools that can fabricate copper interconnects that include cobalt or ruthenium
• Any equipment capable of area selective deposition of a barrier or liner using an organometallic compound.

 

​Tech rag headlines over the last two years have been ripe with a constant stream of semiconductor capacity expansion projects across the globe, with the US competing with South Korea, Japan, Europe, China, India, and APAC to offer up the best possible deals to attract major semiconductor fabs to spend a chunk of those trillions of dollars that have been allocated toward increasing global semiconductor capacity.  While the fabs themselves get most of the headlines, there are hundreds of companies that are essential for those fabs or are necessary parts of those industries that feed semiconductor demand.
With the total funding for September reaching $2.9b we looked a bit more closely at who and what was attracting those dollars.  The folks at semiengineering.com were nice enough to put together a list of funding for start-ups that occurred in September, which we parsed a bit to see  what sub-groups received the most and least funding.
There were 105 start-ups funded in September, totaling $2.978b US, although 19 of the 105 did not disclose the funding amount, with the largest category being materials, a space that rarely gets mentioned unless there is a shortage of a particular material or a rapid rise or fall in the price.  Batteries were 2nd in terms of funding which is not a surprise given the focus on EV, and CPU 3rd, also not surprising with AI being mentioned by almost everyone as a way to differentiate their products, regardless of whether its has real relevance.  While AR/VR, an outgrowth of the popularity of using the word ‘metaverse’, saw relatively small funding on a per company basis, it was tied for 3rd with materials for the most number of companies funded.  The full list of 105 companies and the details on each company can be seen here, thanks to semiengineering.com.

Way back in April of 2019 On Semiconductor (ON) agreed to purchase the Global Foundries (GFS[1]) Fab 10, a fab GF purchased from IBM (IBM) less than four years earlier.  While $100m of the $430m was paid upfront, the balance due will be paid when the deal officially closes at the end of this year.  As ON did not acquire full ownership back in 2019, the two companies reached an agreement where the GF foundry would supply wafer production to ON starting in 2020, while ON was granted rights to GF’s 300mm wafer technology.  When the deal closes ON will produce 300mm wafers for GF through 2025.  As part of the deal Empire State Development provided grants of $17.5m and $22.5m in tax credits toward the purchase that have been and will be paid out over 10 years.
On Semi has offered jobs to what we expect is the 950+ workers directly involved with Fab 10 according to NYS notices, and Global Foundries is expected to offer the remaining ~150 employees employment at the company’s Fab 8 located in Malta, NY.  On Semi has also committed to spending ~$720m over that same 10 year period to enhance the site and hire additional personnel from the region.  While those directly involved in the fab will see little change other than a new name on their paychecks, the GF fab in Malta is ~2 ¼ hours away and would be quite a haul for those that have to make that hike every day.


[1] Global Foundries was formed as part of the AMD divestiture in 2009 and received $1.5b to take over the IBM fabs in Fishkill (Fab 10) and Essex Junction, Vermont (Fab 9).  GF was owned by the Sovereign Wealth Fund of Abu Dhabi until 2021.

​Intel announced a new graphics card aimed at squeezing industry leader NVidia (NVDA) by offering the new card at $329.  Intel is citing the run up in GPU (graphics processing unit) prices, a result of a tight semiconductor market and the needs of bitcoin miners, who use the cards for hash processing, as their focus for the new device, which the company says will have 65% better ray tracing capabilities (image rendering) than the competition, at a more moderate cost, which is a dig at Nvidia’s RTX 30360 graphics card.  The Nvidia card, which appeared on Amazon (AMZN) at ~$480 in May, saw a spike to ~$605 in August and then a fall to $381 earlier this month (currently $399) as demand weakened and inventory levels remained high.
It is early in the lifecycle of the Intel card (Arc A770) so only a few benchmark studies have been done, with the Intel card outperforming the RTX in many, but not all games.  On average the Intel card saw a 13% improvement in frame rate, and Intel says it should see improved performance in those games where it did not beat the RTX as it develops new drivers.  If that is the case, the Arc A770 will be able to compete with the more powerful RTX 3060 Ti card, which retails for ~$450 at least for now as Nvidia and AMD (AMD) are scheduled to release their next generation of GPUs by 1Q ’23.  While Intel has a shot to gain some attention as a lower cost alternative, gamers tend to “Go big or go home”. 

​Apple (AAPL) is considering adding memory products sourced from Yangtze Memory Technologies (state), a Hubei Province based company we have mentioned a number of times, particularly in reference to the company’s efforts to replace as much US made equipment in their semiconductor process fabs as possible with Chinese sourced tools (05/10/21),  our note on the investigation of China’s “Big Fund” management company (07/29/22), which has funded much of China’s semiconductor efforts, and Foxconn’s (2354.TT) attempt to purchase a 10% stake in Tsinghua Unigroup (state), which would give it more direct access to UNISOC (pvt) and YMTC (08/10/22).
While the company is not on the US Commerce Department’s ‘entity’ list, there has been some political discussion concerning the company’s ties to Tsinghua University, a state run institution that is said to have a number of laboratories that are researching defense related projects.  YMTC began producing 64 layer 3D NAND in 2019 and has moved to 128 layer using a 20nm node, but unlike SMIC (688981.CH) and Fujian Jinhua IC (state) the company remains off of the DOC list, with support from a number of global semiconductor associations (SEMI.org, SIA, ITIF, etc.), while a number of think tanks and Congressmen have suggested the company’s ties to the Chinese military should warrant its inclusion under the terms of the 2018 Export Reform Control Act.  Last month rumors surfaced that pressure from Commerce Secretary Gina Raimondo pushed congressional committees to examine the company, China’s largest memory producer, to see whether Congress would support its inclusion to the entities list.
Apple is said to be considering the company’s products for use in some iPhones assembled and sold in China, with no intention of their use in products that leave the Mainland, but members of the House Foreign Affairs Committee have suggested that “Apple will effectively be transferring knowledge and knowhow to YMTC that will supercharge its capabilities and help the Chinese Communist Party achieve its national goals.  While we expect that if Apple were to source NAND from YMTC, there would be some technical communication between the two, especially given Apple’s high quality control standards, but with much of iPhone assembly done in China, Chinese technology spies would already have access to much of Apple’s design technology, and could reverse engineer much of what exists in an iPhone, so the panic concerning Apple sourcing YMTC NAND seems a bit overly dramatic. 
From the perspective of broadly not wanting to help the Chinese semiconductor effort in any way, including the purchase of any components, adding YMTC to the entities list would slow the Chinse semiconductor industry’s growth to a degree, but the politics around anti-China sentiment, especially concerning semiconductors, seems to give politicians a platform that is hard to counter without seeming anti-American or anti-worker.  While there is certainly a very legitimate goal in not supporting the expansion of the Chinese military effort, Apple’s purchase of YMTC NAND flash is going to do little to move the needle in China’s ability to produce ICBMs and other military equipment.  If US politicians are so concerned about the development of China’s semiconductor efforts, perhaps a more unified vote (68 for/32 against) and less resistance to the recently passed CHIPs Bill would have shown China they have a cohesive adversary in the semiconductor space rather than a bunch of self-serving politicians.

  Alabama: Shelby (R-AL), Nay                  Tuberville (R-AL), Nay
 Alaska:      Murkowski (R-AK), Yea           Sullivan (R-AK), Yea
 Arizona:    Kelly (D-AZ), Yea                    Sinema (D-AZ), Yea
 Arkansas:  Boozman (R-AR), Nay              Cotton (R-AR), Nay
 California: Feinstein (D-CA), Yea              Padilla (D-CA), Yea
 Colorado:  Bennet (D-CO), Yea                  Hickenlooper (D-CO), Yea
 Connecticut: Blumenthal (D-CT), Yea        Murphy (D-CT), Yea
 Delaware:  Carper (D-DE), Yea                  Coons (D-DE), Yea
 Florida:      Rubio (R-FL), Nay                              Scott (R-FL), Nay
 Georgia:    Ossoff (D-GA), Yea                  Warnock (D-GA), Yea
 Hawaii:      Hirono (D-HI), Yea                  Schatz (D-HI), Yea
 Idaho:        Crapo (R-ID), Yea                    Risch (R-ID), Yea
 Illinois:       Duckworth (D-IL), Yea            Durbin (D-IL), Yea
 Indiana:     Braun (R-IN), Nay                              Young (R-IN), Yea
 Iowa:         Ernst (R-IA), Nay                     Grassley (R-IA), Yea
 Kansas:      Marshall (R-KS), Nay               Moran (R-KS), Nay
 Kentucky:  McConnell (R-KY), Yea           Paul (R-KY), Nay
 Louisiana: Cassidy (R-LA), Nay                Kennedy (R-LA), Nay
 Maine:       Collins (R-ME), Yea                 King (I-ME), Yea
 Maryland: Cardin (D-MD), Yea                 Van Hollen (D-MD), Yea
 Massachusetts: Markey (D-MA), Yea                  Warren (D-MA), Yea
 Michigan:  Peters (D-MI), Yea                              Stabenow (D-MI), Yea
 Minnesota: Klobuchar (D-MN), Yea           Smith (D-MN), Yea
 Mississippi: Hyde-Smith (R-MS), Nay        Wicker (R-MS), Yea
 Missouri:   Blunt (R-MO), Yea                             Hawley (R-MO), Nay
 Montana:  Daines (R-MT), Yea                  Tester (D-MT), Yea
 Nebraska:  Fischer (R-NE), Nay                 Sasse (R-NE), Yea
 Nevada:     Cortez Masto (D-NV), Yea       Rosen (D-NV), Yea
 New Hampshire: Hassan (D-NH), Yea       Shaheen (D-NH), Yea
 New Jersey: Booker (D-NJ), Yea                Menendez (D-NJ), Yea
 New Mexico: Heinrich (D-NM), Yea           Lujan (D-NM), Yea
 New York: Gillibrand (D-NY), Yea            Schumer (D-NY), Yea
 North Carolina: Burr (R-NC), Nay            Tillis (R-NC), Yea
 North Dakota: Cramer (R-ND), Nay                    Hoeven (R-ND), Nay
 Ohio:          Brown (D-OH), Yea                 Portman (R-OH), Yea
 Oklahoma: Inhofe (R-OK), Nay                 Lankford (R-OK), Nay
 Oregon:     Merkley (D-OR), Yea               Wyden (D-OR), Yea
 Pennsylvania: Casey (D-PA), Yea              Toomey (R-PA), Nay
 Rhode Island: Reed (D-RI), Yea                 Whitehouse (D-RI), Yea
 South Carolina: Graham (R-SC), Yea                  Scott (R-SC), Nay
 South Dakota: Rounds (R-SD), Yea           Thune (R-SD), Nay
 Tennessee: Blackburn (R-TN), Nay             Hagerty (R-TN), Nay
 Texas:        Cornyn (R-TX), Yea                 Cruz (R-TX), Nay
 Utah:         Lee (R-UT), Nay                       Romney (R-UT), Yea
 Vermont:   Leahy (D-VT), Yea                             Sanders (I-VT), Nay
 Virginia:    Kaine (D-VA), Yea                             Warner (D-VA), Yea
 Washington: Cantwell (D-WA), Yea                    Murray (D-WA), Yea
 West Virginia: Capito (R-WV), Yea           Manchin (D-WV), Yea
 Wisconsin: Baldwin (D-WI), Yea                Johnson (R-WI), Nay
 Wyoming: Barrasso (R-WY), Nay              Lummis (R-WY), Nay
 

Semi.org is a global industry association that represents electronics manufacturing and design supply chain covering ~1.3m industry workers and more than 2,500 members.  The organization tracks various aspects of semiconductor industry spending and publishes forecasts and data for its members and occasionally, the public.  The organization just released expectations for semiconductor equipment spending for this year and next, estimating 14.5% growth this year and 2.8% growth in 2023.  Based on their expectations for 2022 the implication would be for 2H sales of $66.39b, up 29.9% over the 1st half and up 22.5% y/y. At 56.5% of the full year estimate, the year is a bit weighted toward 2H as the typical 2H ratio of full year sales is 52.4%.
Almost all of the growth in 2023 will be from Foundry & Logic equipment, which is expected to grow from 55.2b to 59.5b, while memory/storage and Assembly & Packaging tool sales are expected to decline modestly from 7.8B to 7.7b, with Test & Measurement tool sales increasing slightly by 0.4%.  On a regional basis, Taiwan regained 1st place over China in 2Q after falling behind Mainland equipment sales in 4Q of last year, although both regions had almost identical spending in 2Q.  As we have previously noted, Chinese foundries have accelerated purchases of lithography equipment in anticipation of the US tightening restrictions on DUV tools following restrictions on more advanced node EUV tools.  While China continues to expand its semiconductor foundry business at a rapid pace, we expect the value of foundry and locic equipment sales will see less growth as US trade restrictions further limit tool purchases to mature nodes, while Taiwan and Korea are able to purchase those higher priced but higher value tools without restriction.

Over the last few months we have been researching the potential for the advancement of automation in the semiconductor space, meaning the ability of a semiconductor fab to self-correct as silicon moves down the process line, with the ultimate purpose being to increase yield.  As part of the research we have found that such a system, and even systems currently in place, would not be limited by a lack of data necessary to make such corrections in-process, but quite the opposite.  In most fabs, and even in much less complicated production processes, there is too much data, a result of smaller and more accurate sensors that can be placed in more locations with better sensor measurement capabilities and faster data collection networks.
We have found that while there are some data standards that span a number of semiconductor processes, it is the job of process data engineers to look at data generated from each piece of equipment in order to analyze where the line is facing issues, and this could mean data from many sensors on each process tool.  As data collection sensors have improved and collection points increased, the amount of data that needs to be reviewed has become overwhelming, to the point that it can be a hindrance rather than a help.
One vivid example was shown to us concerning a relatively simple process in the CE world, solder reflow inspection on PC boards, a process where the board is pre-heated, solder paste[1] added, and then cooled, which attaches components to the PC board.  While reflow is a basic process that has evolved over the years, typical problems that occur during reflow, such as component shift or delamination of the board, can not only cause the board to be pulled off the line for further inspection (and cost), repair, or to be scrapped entirely, but can become a source of serious problems when the boards reach customers in products.
In many cases PCB boards are inspected visually as they pass through a line, although humans are easily distracted or fatigued, leading to the use of AOI (Automated Optical Inspection) tools that use image sensors to match a board’s captured image to a stored template.  However even AOI tools are subject to errors, particularly false defects, which can be caused by light reflection or other optical interference, which pushes more sophisticated PCB lines to use tools that perform a number of measurements in addition to template matching as the boards pass through the production line. 
One example we saw was a system that was evaluating PCB boards that were ~8.3” x 6.3”, a bit more than half the size of a sheet of paper.  The tool measured height, width, and area at over 3,400 points on the board and checked for x/y offset (components that were placed incorrectly during reflow), generating 17,000 data points for each board.  The data was fed to SPC (Statistical Process Control) software, which, in theory, is supposed to correlate the data, allowing a process engineer to track back where defects were being generated in order to make corrections to the line. 
What was most surprising was that in many cases, when a board was flagged by the system, the process engineer let the board remain in the line, under the theory that the final AOI and SPC review would pull the board before it was shipped.  This defeated the purpose of the entire system as the flagged board continued to be processed, despite the fact that it had defects and would either need repair which could entail removing later process steps, or could be scrapped entirely, and if the defect was not clarified and the reflow process fixed, it could continue to affect subsequent boards.   When asked why the engineer let the flagged board go through, the comment was that there are so many data points that the chance for false negatives continues to increase with the number of measurement points, so rather than pull the board they left the determination up to the final inspection engineer, essentially defeating the purpose of the data collection early in the process line.
While this was an example using a relatively simple process and tool set, producing semiconductors is far more complex, with memory producer Micron (MU) stating that it takes more than a thousand process and measurement steps to produce modern chips.  To put just the feature size of those structures into perspective, think of a chip die, typically about the size of a fingernail, being blown up to the size of a football field.  Take one blade of grass from that field, cut it in half, cut it in half again, and cut it in half a third time.  That’s the size of one transistor, essentially one bit of storage, out of the 8 billion on a typical memory chip, so the amount of data generated by tools to make sure things are where they should be on semiconductors is enormous, more than all the data collected by Facebook (FB), Amazon (AMZN) and all the other social media companies, although over 80% of it is not ever looked at by process engineers.
Most semiconductor process engineers are only interested in data when there is an ‘excursion’ in the semiconductor fabrication process, a nice way of saying ‘an error’, as these can cause millions in losses considering the number of die on each wafer and the number of wafers processed each hour.  Some semiconductor customers require (military, automotive) that the data be stored for up to 15 years, even if it has never been reviewed, but ~90% of the available semiconductor process data is less than two years old according to T&M companies, which means the amount of data being generated during the last few years has increased dramatically.
So the question then becomes whether it is actually helpful to add sensors, data collection points, and sample rates if most of the data is never going to be used?  The fear that at some point the data might be useful pushes the industry to collect everything, even if it seems unnecessary, ‘just in case’, but does it really help to improve yield, the ultimate goal, if most of the data is ignored?  Not really, so improving sensors, adding data collection points, and increasing data collection network speeds all sound good on paper but the increases in process data volumes don’t seem to lead to large gains in productivity and yield, so where is the problem? 
The problem is that humans are not interested in the data as a whole, only if it solves an immediate problem, while computers are designed to process data, especially large amounts of data, and can be ‘trained’ to spot even the smallest anomaly.  This is essentially what SPC software is supposed to do, help a human engineer spot a problem and  trace it back to its source, a process that even in the PCB example above, would entail the engineer having to look at the data from a number of tools in order to spot the root cause of the error, something that computers can do far more easily than humans, and if we trusted the computer, we could actually let the computer make the necessary adjustments to the process tools without getting involved.
This is the goal of semiconductor foundries; to fully automate the process using a self-correcting system that could make corrections on a tool-by-tool or process by-process-basis, leaving the human engineering staff to come up with better and more efficient process steps.  Unfortunately there is a problem, and that comes back to the data.  Much of that data is collected at the tool level and tool manufacturers want users to be able to use that data through a platform specific to that tool while the next tool vendor wants the user to use his platform for the data, leading to lots of data that is not uniform, especially as tool vendors and the industry overall is just coming to the conclusion that the amount of data is so vast that AI and machine learning systems are the only way in which the data can be utilized.  This leads to the conclusion that the data must be uniform and organized to avoid the need to ‘clean’ the data or reformat same before it can be processed, which puts bottlenecks into the potential feedback loop.
There are some tools that are self-correcting, such as ASML’s (ASML) EUV tools that allows the tool to measure and compensate for sub-nanometer inaccuracies that occur as the scanner operates, feeding back that measurement data directly to actuators at the reticle stage, which compensate for those minute changes that would otherwise become potential excursions.  ASML is the first to admit that the stresses that their systems work under, such as the fact that the system must move the wafer position to within a quarter of a nanometer 20,000 times each second,, must be checked, compensated for and adjusted with each movement for the system to work correctly.  In such a system there is no human intervention, which forces engineers to trust that the data is being correctly analyzed and the compensation increases the tools yield.
That said, the goal of a unified data set and a master AI system for a semiconductor fab that could be self-correcting is a lofty and far-off goal, but the challenge is more to bring uniformity to fab data and reduce the number of proprietary data ‘silos’ that exist in fabs currently.  It is an enormous task, but one that must be addressed as the complexity of semiconductor processing continues to increase and as the data volumes continue to increase, we expect fab spending for software systems designed to work with disparate systems will increase, while the industry reaches for standards that will shift that focus from data compensation and presentation to generating AI systems that are able to look across a fab, find production bottlenecks and compensate for this issues by adjusting tool parameters anywhere in the fab.  Might not be in our lifetime, but we will keep digging in coming weeks to see if anyone is on track to making a dent in all that data and whether there is any hope of a unified solution that could make the vision of an automated fab a reality.
Note:
FOUP – Front Opening Unified Pod – Wafer Carrier
FDC – Fault Detection & Classification
EUV – Extreme Ultraviolet – Lithography tool
 


[1] Solder paste is a mixture of solder powder and a flux.  The flux holds the components in place, prevents oxidation, and allows the melted solder to flow.  The solder itself completes the connection between the component and the traces (think of flat wires) on the board.

As Russia continues it war against the Ukraine, a number of Chinese manufacturers have been forced into the spotlight with their products being mentioned by Russian military authorities, or being shown in compromising situations in the media.  Last month SMIC (688981.CH), China’s largest semiconductor foundry, answered investor questions about it ties with Russia stating that it “ …has never had customers in Russia, and remains in compliance”, in order to keep from generating more ill will with the US government, as the company is already on the trade restriction list limiting its ability to purchase advanced semiconductor tools and materials.. 
Chinese drone producer DJI (pvt) also has spent time trying to allay fears that its drones are being used by the Russian military after a statement of praise by a Russian official began circulating in social media concerning the company’s products.  Now, China’s Unitree Robotics (pvt), a producer of a $2,700 robotic ‘dog’, is trying to disarm the chatter that occurred when one of its robotic dogs was seen at a military hardware exhibition in Russia with a rocket launcher attached.  While the robotic dog was covered in black cloth, it was quite similar to the company’s household robot dog, the GO1, and while the Russian engineers who developed their implementation said it can be used in both civilian and wartime scenarios to deliver medications, they also noted it could be used to carry and fire weapons.
Boston Dynamics (pvt) has been building commercial quadruped robots since 2004, brothers of the dancing robots we have seen in a number of demos and segments on 20/20, but there is always the inevitable corruption of such devices for use in the military, singled out in a number of dystopian Sci-Fi thrillers that feature ‘Spots’ with weaponry and a mindless focus on killing hapless civilians.  While companies are very careful to disassociate themselves from the current worldwide bad guys and are hopefully careful not to sell directly to any military organization (other than our own) that might adapt them for warfare, it would be hard not to imagine that the human mind would not look at such devices as potential weapons, especially if they are able to be reverse engineered and easily copied.  Covering them with a black cloth (as shown in the first video below) does little to hide their origin and sets a depressing tone to what are certainly feats of engineering, so we also include the second video to put things in a better light…
https://youtu.be/WZlMq5LpN8Q
https://youtu.be/fn3KWM1kuAw
​