Sunday, April 30, 2023

Introduction to Semiconductors

 An overview of the semiconductor industry.  Aims to understand:

  • What are the moats in the industry?  
  • When you are looking at a company, where does it sit in the ecosystem?
  • Long term trends.  Trends in this industry take years of industry-wide cooperation and billions of dollars to play out.  Once you can see a trend start, you can pick tomorrow's winners.

Types of Chips

Memory Chips

Memory chips were a commodity subject to wild supply and demand swings. Now the industry has consolidated into 3 players: Samsung, SK Hynix (both Korean) and Micron.

Source: For the first time in Memory, have durable profits finally arrived?

China's YMTC recently entered the market, aiming for a 6-8% market share in 2022.  Their chips are state-of-the-art, but YMTC is on the US blacklist.

Despite its oligopoly structure, the memory industry is in a downturn (crash?) now due to falling demand from a post-covid hangover.  While Samsung is maintaining capex.  The memory industry is still irrational.

In the long run:

  • The Koreans look like winners due to scale and predatory capex/pricing.
  • Micron looks like a loser due to debt and lack of EUV.
  • YMTC will operate in its own market, but will slowly fall behind from lack of EUV.

Logic Chips

Two main types:

  • CPU's run general instructions to allow a computer to do anything.  They are designed by Intel/AMD (for computers), or ARM (originally for handphones - now also computers).
  • GPUs are specialised chips to perform simple math calculations.  Used for graphics displays (eg: realistic games) and AI (matrix calculations).  The leading GPU designer is NVDIA - they have a moat from their CUDA interface being a standard, but its being challenged.

Logic Chips are built in Fabs, usually separate from the company that designed them.  The foundry industry has consolidated so that only three companies can build modern logic chips:

Of the three leading-edge players:

  • Samsung is fucked for high end logic chips.  They have a top down management culture unsuited to technology development and are serial liars.
  • TSMC is the clear leader due to their partnership with Apple.  The risk for TSMC is a Chinese blockade/invasion of Taiwan - I give it a 30% chance before 2030.  Even without war, new foundries have to be built elsewhere, which has to be bad for the market in the long term.
  • Intel has historically fabbed its own chips from when it was the leader in the 90's, but has now fallen behind both in chip design and foundry capability.  Intel has a bad capital structure and was bleeding to death.  They may have stopped the bleeding with their Feb dividend cut.

CPUs are currently in a downturn due to the covid-hangover, but not as bad as memory.  

GPUs (including ASICS for crypto mining) are in a downturn, but NVDIA's stock is in an AI bubble.

Analog

Analog chips are simple ones used to convert "real-world data" (eg: light/voice for camera/phone) into digital signals to processed by other chips.

Players are Texas Instruments (design and fabs), Analog Devices (design only) and Tower Semiconductor (fabs).  All these companies look interesting.  Haven't looked at their valuations or the industry cycle.

System on a Chip (SOCs)

Up until 10 years ago, we made computers faster by improving their CPUs: squeezing more transistors into a piece of silicon, and increasing the clock speed (time to process one instruction).

This stopped working, and we now combine CPU, GPU and memory onto a single chip.  The idea is to allow different parts of the chip to perform specialised tasks (eg: in your phone: a GPU for video processing, with a less powerful CPU for general processing).  It allows more tasks to be done faster with less processing and heat.  Apple's desktop M1 chip used this in 2020 to leap over previous generations in terms of performance.

Some consequences of this are:
  • Expect the decline of Intel/AMD chip designs in desktop/server computing.  Intel/AMD's x86 or x64 chips use CISC (complex instruction sets) which are unsuitable for SOCs.  The only advantage of x86/x64 is that you can run legacy DOS/windows programs developed since the early 80's.  ARM chips (which are non-x86 or RISC) were not able to run them, but this has changed.  Apple now has an emulator to run intel-compiled programs.  Microsoft is developing an arm based server and tablet devices.  Expect Wintel to die.  Expect the majority of servers/desktops to move to ARM based processors, while x86/x64 chips become a specialised market for running legacy software.
  • Big Tech (Apple, Google, Microsoft, Amazon) design their own in-house chips for their products/operations.  e.g.: Amazon's Nitro SOC runs the hypervisor (VM manager) in their AWS, freeing up resources for user operations.  Google's TPUs run AI calculations.  This vertical integration can be a way to extend their moats.  (...Or maybe over-optimisation just distracts them from their real business: AmazonGoogle).
  • Increasing importance of the "packaging" part of the chip production process (below).

Classifying Chips by Usage and Size

Another way to categorise chips is by size (nm).  When discussing chips, you need to place them by size and timeline to know what parts of the industry you are talking about. 

Roughly they can be broken down into:

  • High End: For phones, PCs and Servers.  Produced in Taiwan, the US and South Korea.  Designed in the US.  Probably anything smaller than 28nm.
  • Medium End:  Probably 28nm or above.  For automobiles, aerospace.  Produced in Malaysia, Thailand, Philippines, Singapore, plus the above countries.
  • Low End: For the IOT.  To help your refrigerator keep track of your shopping list.  Produced in China.
This diagram shows it visually:

There was a huge shortage of medium end chips for car manufacturers in 2021.  Foundries cannot easily add capacity for these older chips, which are sold so cheaply because their production facilities are fully depreciated. New production facilities require higher prices and long term agreements.

Node sizes for the 3 foundry businesses are not equivalent: Intel's 10nm is equivalent to TSMC's 7nm (so its now called "Intel 7"), and Intel's 7nm ("Intel 4") is equivalent to TSMC's 4nm.  See "Node Name Disconnect".

Production Process

The 5 steps of chip production.  Generally, the outputs of each step are the inputs to the next:

Wafers

Produce the silicon wafers that chips will be printed on.  The wafer market was an oligopoly, but may now be cracking:

  • 5 companies produce(d?) 90% of silicon wafers (p5).  2 of those are Japanese: Sumco and Shin Etsu.
  • 150mm and 200mm wafers are for older chips, 350mm are for newer ones.  A proposal for 450mm wafers was killed by TSMC in 2013.  The industry hasn't changed much since the 90's.
  • Sumco expects no capacity expansion till 2024, despite a wafer shortage.
  • This may give space for Chinese companies to enter the market, especially for 150 and 200mm.  Chinese firms expanded production in 2021.  South Korean imports form China almost doubled from 2020 to 2022.  Chinese wafers are reportedly 5-10x cheaper than Japanese ones.

Lithography (ASML)

Lithography is the process of etching circuits into a wafer.  A good description is in Doug O'Lauglin's Fabricated Knowledge.  Chips 7nm or below require EUV, which is from one company: ASML.

ASML's EUV machines are incredibly complex and precise:
  • "molten tin droplets of around 25 microns in diameter are ejected from a generator at 70 meters per second. As they fall, the droplets are hit first by a low-intensity laser pulse that flattens them into a pancake shape. Then a more powerful laser pulse vaporizes the flattened droplet to create a plasma that emits EUV light. To produce enough light to manufacture microchips, this process is repeated 50,000 times every second."  
  • ASML's supply chain is worldwide.  Their suppliers are small companies that each specialise in one out of the hundreds of steps required. (eg: A small company in Vienna selling equipment to produce photomasks).
No one can replicate this and integrate the work of hundreds of suppliers to the required precision.  Especially when the market is limited to three customers who all invested in ASML in the first place.  ASML is a monopoly for the next ten years.

If I was buying ASML, valuation is the only question and the key number is their recurring revenue for their installed base vs new sales.  The stock is too expensive now.  ASML or its suppliers may be a cyclical play later.

Applied Materials Sculpta process may halve the need for EUV.  Or maybe not.  Hard for lay people to judge.

Inspection

Covers non-intrusive ays to measure a chip to make sure its correct.  Theoretically it means using beams of electrons, x-rays or light to measure the surface of a chip.  The complexity is in interpreting information from the reflected beams to form a picture of reality.  More details from Fabricated Knowledge's Meterology Primer.

Packaging

Attaching the silicon chip onto a motherboard. From old to new:
  • From the old days, a DIP chip package, with its iconic "spider legs":
  • Wire bonding: Using copper wire to attach the chip to the circuit board.
  • Flip Chip: Depositing an array of solder balls onto the board's pads, and "flipping" the chip onto it.
  • 2.5D packaging: TSVs passing through the wafer (like underground cables) connect different chips on the board.  Used for SOCs.
  • 3D packaging: Connect different chips which are all sitting on top of each other (on the circuit board).  Currently used in memory chips, not yet in logic chips.

Testing

Theres two levels of testing:
  • Functional Testing: Running a current through a chip to test input and output combinations.
  • System Level Testing: Test the chips under the conditions they are expected to be used.  eg: Test at varying temperatures, or for long run-times (burn-in), structural tests, or software level testing (calling the chips functions/libraries in the same way that software wold).
The big companies involved in the last 3 steps are Applied Materials, Lam Research and KLA Tencor.  

2.5D or 3D packaging may be done more by the fabs (TSMC). 

Conclusions

Theres less than 10 big players forming the semiconductor ecosystem.

Things change slowly, but they do change.  eg: Wintel's fall and TSMC's rise.  These are not Buffet-like stocks to pass to your grandkids.

I believe China will not be able to make high end logic chips - they would have to replicate an entire ecosystem.  They can do high end memory.

Biggest moat is for ASML.  But everybody knows it and they are expensive.

TSMC is the clear foundry leader, and fabs have pricing power.  They are reasonably priced, though probably its not the right time in the cycle.  Big geo-political risk.

Intel was dying.  They have stopped the bleeding.  Maybe a turnaround play?  Can they get the money to build Fabs in a downturn, and as the market moves away from CISC?

Resources