Who Actually Makes a Semiconductor? Mapping the Companies Behind Every Step from Design to Board Assembly

Open on YouTube ↗
Overview

Electronics creator Ichiken built this video from study notes they compiled while learning about the semiconductor industry. They found that many books and materials explain how chips are manufactured, but few let you see which companies are involved at each step. The video follows a single chip from design through materials, masks, wafers, front-end and back-end processing, testing, and finally mounting on a circuit board, and names the equipment makers, component suppliers, and materials companies active at each stage.

34 min read

The focus is on cutting-edge logic semiconductors: the chips that do the computation in smartphones, PCs, and AI servers. Memory and power semiconductors come up along the way, but the processes and companies were chosen around the workflow for advanced logic. Ichiken's stated goal is that when a semiconductor company appears in the news, viewers can tell where in the process it sits.

2:29

The Overall Flow and Where Companies Fit

Ichiken starts with a process map. Three things must be prepared before manufacturing can begin: the circuit design, the photomask that carries the pattern from that design, and the wafer that forms the base of the chip. The central section is where chips are actually built. Circuits are formed on the wafer in the front-end process. While the chips are still on the wafer, they are tested and the good ones are identified. The wafer is then cut apart in the back-end process, the chips are packaged, and a final test is run before shipment. So there are two tests: one before dicing and one after the chip is finished as an electronic component.

Three types of chip-making companies cover this section: foundries, which manufacture other companies' designs; IDMs, which do everything from design to manufacturing in-house; and OSATs, which handle assembly and testing. Finished chips are then mounted on products during PCB assembly. Underneath all of this sit companies that people tend to overlook: suppliers of materials, of components that go inside the manufacturing equipment, and of factory infrastructure.

Ichiken's overview slide arranges company types from left to right as design, wafers, masks, front-end, back-end, testing, and board assembly, with Japanese companies highlighted in red. Laid out this way, Japanese firms cluster in front-end and back-end manufacturing equipment, in the components used inside that equipment, and especially in materials. The bottom rows show the chipmakers themselves (foundries, memory makers, IDMs) and the infrastructure and component suppliers that support the whole process.

For deeper reading, Ichiken recommends two resources: slides on the semiconductor industry compiled by "🐇の半導体勉強部屋" (🐇の半導体勉強部屋), which list companies by sector, from EDA and IP through wafer equipment, materials, gases, chemicals, and trading companies; and a process explanation on semiconductor-industry.com by a working process engineer known as ずーぼ, which walks through the front-end and back-end step by step.

5:18

Two Business Models: Horizontal Division of Labor vs. Vertical Integration

There are two broad ways to make semiconductors. The first is a horizontal division of labor, in which design, manufacturing, and assembly/test are handled by separate companies.

  • Fabless companies only design. Examples are NVIDIA and Apple, and recently Google, Amazon, and Microsoft, companies whose core business isn't semiconductors but which design chips for their own use.
  • Foundries handle manufacturing, specifically the front-end process. TSMC is the best known. Rapidus, which aims to mass-produce cutting-edge logic in Japan, is also a foundry, as is Samsung. Ichiken notes that while foundries aren't household names like their fabless customers, most people have heard of them in the news.
  • OSATs handle the back-end and testing.

The second model is vertical integration: the IDM. Ichiken lists Intel, Texas Instruments, and Renesas; power semiconductor makers Infineon, ROHM, Mitsubishi Electric, and Fuji Electric; and Sony for image sensors. Memory makers such as Samsung, SK Hynix, Micron, and Kioxia also design in-house and manufacture in their own fabs, so Ichiken treats them as close to IDMs. Samsung and Intel are IDMs that also run foundry businesses for other companies.

Ichiken stresses that the two-way split in the diagram is a simplification. Many companies sit in between. A fab-lite company manufactures in its own fabs but outsources only its most advanced products to companies like TSMC. Infineon outsources some microcontroller production to foundries, and Sony outsources part of its image sensor production. Most companies decide product by product how much to make in-house, so the slide simply groups companies by which side they lean toward.

7:50

Key Players: Fabless, Foundries, IDMs, and Memory

Among fabless companies, Ichiken observes that each has its specialty: NVIDIA in GPUs, AMD in CPUs, Qualcomm in smartphone SoCs (Snapdragon), and Broadcom in custom designs such as Google's TPU, an AI-specific computing chip.

For foundries, TSMC is the name heard most often, and for cutting-edge logic there are products that seem only TSMC can make. Others include Samsung, Intel Foundry, SMIC, and Rapidus. Ichiken says they have high hopes for Rapidus, which is working hard to manufacture 2nm-equivalent chips in Hokkaido.

Among IDMs, Intel makes PC and server CPUs in its own fabs and has recently moved into contract manufacturing. Sony leads the world in camera image sensors. Renesas is a major maker of automotive microcontrollers. Infineon is a leader in power semiconductors; besides standard silicon devices, it makes power semiconductors using silicon carbide (SiC) and gallium nitride (GaN), which tolerate high voltages and fast switching. Texas Instruments leads in analog ICs for power, audio, and sensor signals. Mitsubishi Electric, Fuji Electric, and ROHM are Japanese power semiconductor makers, each with its own strengths. Some companies on the list have fab-lite elements, but Ichiken grouped them as IDMs for this video.

Memory makers are a prime example of IDMs. Ichiken explains the two main types. DRAM temporarily holds data being processed and loses it when power is cut. HBM, the high-bandwidth memory that has been in the spotlight for AI, is a type of DRAM that stacks multiple DRAM dies so large amounts of data can be exchanged at once. NAND stores photos and large data and retains it without power. The biggest players are concentrated in South Korea, Samsung and SK Hynix, which Ichiken thinks look set to be among the winners of the AI era. Micron is another well-known name, and Kioxia in Japan makes NAND. Both DRAM and NAND are increasingly built by stacking layers to raise speed and capacity. Chinese companies such as CXMT and YMTC have also been gaining ground.

11:20

Design: EDA, IP, and the PDK

Ichiken first clarifies two terms. EDA is design software, and IP refers to pre-designed circuit blocks. A modern chip contains tens of billions of transistors and can't be designed by hand, so designers use EDA tools. Circuits with well-defined functions, such as CPU cores or USB interfaces, are bought as ready-made IP and combined on the chip. Fabless companies like NVIDIA and Apple and IDMs like Intel and Renesas both use EDA and IP. The difference lies in where the chip is manufactured.

That difference shows up in the PDK (Process Design Kit), which Ichiken describes as a collection of design rules for a particular fab. Because the shapes and sizes that can be manufactured differ by fab, once you choose a fab you design to its rules. A fabless company receives a PDK from a foundry such as TSMC, designs to it, and outsources manufacturing. An IDM uses its own PDK and builds the chip itself.

In EDA, the three leaders are Synopsys, Cadence, and Siemens. Synopsys offers a full suite of core design tools and is also a major supplier of interface IP such as USB and PCIe. It has been acquiring Ansys, a physics simulation company. Ichiken's interpretation of the motive: when chips are stacked, as in AI chips, heat builds up, boards warp, and signals have a harder time getting through. Signals at very high frequencies also interfere with each other, and all of these effects influence one another, so circuit design now has to be combined with physical analysis. Cadence is strong in analog and package design and can handle the package substrate together with the chip. Siemens' Calibre, a final pre-production verification tool, is widely used as an industry standard. Other names include Keysight in the US, which is strong in high-frequency design, and companies doing component simulation. For PCB rather than chip design, there are Zuken and Altium, which is owned by Renesas. China is also developing its own EDA tools.

As for IP, designing blocks like CPUs and USB from scratch for every chip would take too much time and manpower and add risk, so companies buy proven blocks and combine them with their own circuits. IP isn't cheap: Ichiken says there is generally an upfront license fee plus a royalty on every chip sold. ARM is the leading IP vendor. It doesn't make chips; it licenses CPU designs and earns from royalties, a business model Ichiken finds interesting. They add that by 2026 ARM appears to be making its own chips for data centers for the first time. Synopsys and Cadence are also major IP vendors. RISC-V, a CPU instruction set anyone can use for free, has been spreading, with companies such as SiFive in the US and Andes in Taiwan active in that area. There are also IP firms specializing in GPUs, memory interfaces, and on-chip interconnect.

15:45

Photomasks: Expensive Masters for Every Layer

Once the design is done, manufacturing needs two things: a photomask, a glass master plate carrying the circuit pattern, and a silicon wafer to build on.

At the end of design, the layout data is sent to the mask shop, a handoff the industry calls "tape-out." The mask shop transfers the pattern onto glass to make the photomask. During the exposure step of the front-end process, light passes through the mask onto the wafer to print the circuit, and exposure is repeated many times. Because each layer has a different layout, each layer needs its own mask. According to Ichiken, a state-of-the-art product needs roughly 70 to over 100 masks, and a full mask set for single-digit-nanometer chips costs $10–20 million or more, or over a billion to several billion yen. Because the mask is the master, a single scratch is reproduced on every chip exposed through it. A design error means remaking this costly set, which is why a failed design is so expensive.

Ichiken notes that while mask makers make masks, cutting-edge masks are usually produced in-house by chipmakers such as TSMC, Samsung, and Intel. Merchant mask makers like DNP in Japan sell to outside customers.

The process starts with special glass that doesn't expand or contract. A light-blocking layer is added to make "mask blanks," supplied by companies such as HOYA and AGC. An electron beam writes the circuit onto the blank, and there are equipment makers for e-beam writers. Development, etching, and cleaning steps are not shown on the slide. Inspection follows. For EUV masks, inspection uses an EUV light source, and some companies make those sources in-house; Ushio and Gigaphoton (part of the Komatsu group) also produce EUV sources for mask inspection. EUV is the light used to expose advanced logic, and its short wavelength allows extremely fine circuits. Finally, a very thin film called a pellicle is mounted as a dust cover. Because there is a gap between the pellicle and the mask, dust that lands on the pellicle has little effect on exposure.

20:11

Silicon Wafers: From Quartz Rock to 11N Purity

The wafer starts as ultra-high-purity polysilicon, made by companies such as Wacker, Hemlock, and Tokuyama. Polysilicon is melted and grown into a large single crystal, sliced thin, and polished to a mirror finish. Shin-Etsu Chemical is the leading wafer maker, strong in low crystal defect density and high flatness; Ichiken describes it as almost a general materials company, since it also supplies photoresists and mask materials. SUMCO specializes in wafers and focuses on cutting-edge 300mm products. GlobalWafers in Taiwan is another major supplier.

What makes silicon wafers remarkable, Ichiken says, is purity: about eleven nines, known as "11N." Ichiken also gives rough chips-per-wafer figures: around 500 for a smartphone SoC and around 60–80 for a high-end GPU. Defects reduce the usable count further. That is why larger chips cost more, and why chiplets, which build small dies and combine them to act as one chip, have been gaining attention.

The supply chain begins earlier. Quartz or silica rock, made of silicon and oxygen, is heated with carbon to remove the oxygen, yielding metallurgical silicon of about 99% purity, which is far from enough for semiconductors. A further, somewhat tricky process turns it into polysilicon of about 11N. The wafer maker's work then starts: polysilicon is melted in a quartz crucible, a seed crystal is dipped in and slowly pulled up while rotating (the Czochralski method), and a large single-crystal ingot grows. The ingot is trimmed, sliced with thin wire, chamfered to keep the edges from chipping, ground on both sides to even out thickness, chemically etched to remove grinding damage, and polished to a mirror. It is then cleaned, inspected for particles, scratches, and flatness, and packed in special shipping containers. Ichiken stresses that wafer quality depends not only on ingot growth but on the precision of the cutting, grinding, and polishing equipment and materials, and shows a slide of the many companies involved. They find it interesting that even the shipping containers have specialist makers.

In a short digression on non-silicon wafers, Ichiken mentions SiC for EV motor-drive power converters. US company Wolfspeed handles everything from substrates to devices, though its finances in 2025 apparently weren't good. Other examples are InP substrates for optical communications, GaN substrates for high-frequency and optical devices, and GaAs substrates.

24:30

The Front-End Process: Building Layers in a Loop

The front-end builds the circuit as many layers on the wafer. It has the most steps and the largest number of equipment, component, and materials companies.

Ichiken shows a chip cross-section. At the bottom is the silicon wafer, on whose surface transistors are built: billions to tens of billions on the latest chips. Making the transistors is the first half, called FEOL (front-end of line), whose main steps are thermal oxidation, lithography, etching, ion implantation, and heat treatment. Stacked above is the copper wiring that connects the transistors: about a dozen layers, getting thicker toward the top. This is BEOL (back-end of line), using deposition, lithography, etching, plating, and planarization. Cleaning and inspection are inserted throughout.

One layer is formed as follows. Clean the wafer surface. Deposit a film. In lithography, coat resist, expose the pattern, and develop it. Then either etch to remove film, which creates shapes, or implant ions to introduce impurities, which changes silicon's properties; the choice depends on the layer. Strip the resist, apply heat treatment (for example after implantation) to restore the atomic arrangement, polish uneven surfaces flat, and inspect for defects and misalignment. Then start again with cleaning for the next layer.

For each step, Ichiken looks at three kinds of companies: equipment makers, whose machines go into foundry and IDM fabs; component makers, which supply lenses, power supplies, valves, wafer stages, and so on to equipment makers; and materials makers, which sell chemicals, gases, photoresists, polishing slurries, and other consumables to fabs.

27:44

Cleaning, Deposition, and Lithography

Cleaning removes contaminants using chemical solutions and ultrapure water. SCREEN is the world leader in equipment, with Tokyo Electron, Lam Research, and SEMES also active. Because the tools contain liquids, they need seal materials that withstand chemicals, and they contain wafer-transport robots, which are found in most front-end tools. Chemicals such as hydrogen peroxide and hydrofluoric acid (which dissolves oxide films) are an area where Japan has been strong, and there are makers of ultrapure water systems for rinsing. Since keeping the wafer as clean as possible is essential, Ichiken calls cleaning a very important step.

Deposition lays down films in nanometer increments using CVD and ALD (chemical reactions of gases), sputtering (vaporizing metal and depositing it), and thermal oxidation. Applied Materials, a US company, has the widest lineup, with ASM in the Netherlands and Japanese companies also active. Components include ceramic heaters that warm the wafer, mass flow meters for process gases, and plasma power supplies. Materials, such as sputtering targets for metal films and precursor gases, are also important.

Lithography begins with coating and developing: a thin layer of photoresist is applied, and after exposure, unwanted resist is dissolved with developer to reveal the pattern. Tokyo Electron leads in "coater/developer" equipment, which is used linked to the exposure tool. Materials loom large here, and Japanese companies are particularly strong in photoresists: Tokyo Ohka Kogyo and JSR develop EUV resists, and Shin-Etsu Chemical supplies resists for argon fluoride (ArF) laser exposure, the previous-generation technology. Components include filters that remove tiny particles from the resist and pumps that dispense precise volumes.

Exposure, which Ichiken calls a core step, uses a lens to shrink the mask pattern onto the resist. In state-of-the-art EUV, glass absorbs the light, so mirrors are used instead of lenses. Shorter wavelengths allow finer patterns, and finer patterns mean more transistors per chip. ASML in the Netherlands is the only company that can make EUV lithography systems. Nikon and Canon also make exposure tools, but they cannot produce patterns as fine as ASML's EUV.

The EUV light source is made by Cymer, an ASML subsidiary. Ichiken highlights the "droplet generator," which reportedly fires tin droplets a few tens of micrometers across at tens of thousands per second. A CO₂ gas laser with a wavelength of 10.6 micrometers hits the tin. That wavelength is far too long to pattern fine circuits directly. Instead, the tin turns into plasma and emits EUV light at 13.5nm, which is reflected off the photomask onto the wafer's resist through a highly complex optical system. Previous-generation DUV lithography is still widely used, and for its ArF and KrF laser sources, Japan's Gigaphoton is a manufacturer on par with Cymer. The lenses and mirrors in ASML's EUV tools are made by Zeiss in Germany, and the mirror bases use special low-thermal-expansion glass such as Corning's. Materials include the mask blanks and pellicles mentioned earlier and gases for the light source.

33:54

Etching, Ion Implantation, Heat Treatment, Planarization, and Inspection

Etching uses plasma gas to remove film in areas the resist doesn't protect. Lam Research leads, with Tokyo Electron, Hitachi High-Tech, and Applied Materials also present. Ichiken explains the electrostatic chuck: because the chamber runs at low pressure, air suction can't hold the wafer, and mechanical clamps apply force to the wafer, which isn't ideal, so static electricity is used instead. Other components include RF power supplies for plasma, vacuum valves, and electrode rings. Materials include fluorine- and chlorine-based gases.

Ion implantation ionizes impurities such as boron and phosphorus and drives them into the wafer to give transistors their switching properties. Equipment makers include Applied Materials, Axcelis, and Sumitomo Heavy Industries; Ichiken remarks on how often Applied Materials comes up. A high vacuum is needed to send ions in a straight line, so vacuum pumps are essential. Dopants are supplied as gases, some highly toxic, so storage containers matter.

Heat treatment restores the atomic arrangement disrupted by implantation so the implanted impurities can function in the transistor. Methods range from slow heating in a furnace to heating for just milliseconds.

Planarization (CMP) polishes the wafer flat using both chemical and mechanical action with a slurry and a pad. Applied Materials is strong in equipment, and Ebara is also a player. This step consumes a lot of materials: slurries from companies such as Fujifilm, Entegris, and Resonac, plus polishing pads and pad conditioners. Ichiken admits they didn't know conditioners existed until they looked it up.

Inspection and metrology checks for particles and defects after each step and measures line widths and overlay misalignment. KLA in the US and Hitachi High-Tech in Japan make CD-SEMs, electron microscopes that measure circuit line widths, and ASML also makes overlay measurement tools. Measuring at nanometer scale requires moving the wafer precisely, so THK's linear guides and position sensors matter. Ichiken notes that because this step only measures, they couldn't find any consumable materials for it.

Cleaning, deposition, lithography, etching, implantation, heat treatment, planarization, and inspection make one loop and one layer. Because a chip has dozens of layers, the loop repeats many times. For state-of-the-art chips, lithography alone is said to be repeated from several dozen to around 100 times, which is why a wafer takes several months to complete the front-end. At the end comes the WAT (Wafer Acceptance Test): probes contact test patterns built onto the wafer and current is applied to confirm the transistors were made as intended. Interim inspections check physical form; WAT is the final electrical check of the front-end.

38:24

Shared Components, Factory Infrastructure, and the "Big Five"

Some components appear across many processes. In tools that run at atmospheric pressure (cleaning, coat/develop, CMP) there are transport robots, seal materials, chillers for temperature control, gas flow meters, high-purity valves and fittings that switch gas without contaminating it, and gas panels that bundle the piping. In vacuum or plasma tools (deposition, etching, implantation, EUV lithography, e-beam inspection) there are electrostatic chucks, an area where Japanese ceramics makers are pushing hard; vacuum pumps, vacuum valves separating atmosphere from vacuum, and vacuum gauges; abatement systems that neutralize process gases; RF power supplies; ceramic chamber parts such as walls and showerheads; cleaning and refurbishment services for dirty parts; and quartz parts for heat treatment and deposition.

Factory infrastructure matters too. Wafer containers are carried automatically on ceiling rails, an area where Murata Machinery and Daifuku are strong. The containers are heavy and the volume is huge, so automation is essential. Fabs also need ultrapure water, sealed wafer carriers, and cleanrooms. Ichiken's personal view is that component suppliers are in a good position: whichever equipment maker's tools sell well, they get work.

Summarizing front-end equipment, Ichiken identifies about five major makers. ASML is the only EUV lithography maker. Applied Materials covers deposition, polishing, and implantation and is, in Ichiken's view, probably the broadest lineup. Lam Research leads in etching and excels at digging deep holes for 3D NAND. Tokyo Electron makes the coater/developers used alongside the leading EUV systems. KLA leads in defect inspection. Ichiken adds SCREEN in cleaning, and observes that even the majors have neatly separated areas of expertise.

41:43

The Back-End Process: What Gets Built and How It Flows

The back-end cuts the wafer into chips, mounts and connects them, and seals them in resin. Ichiken says people tend to see these as mundane steps, but AI chips have changed them significantly.

In a package cross-section, the chip from the front-end sits in the middle, connected to the package substrate by thin metal wires or bumps (small solder balls). The package substrate fans out the chip's fine terminals to larger terminals for mounting on the product's PCB. Resin encases everything against moisture and impact, and solder balls on the bottom attach it to the board, as on smartphone and PC boards.

Most cutting-edge chips, including smartphone SoCs, PC CPUs, and AI GPUs, use flip-chip packaging instead of wire bonding. The chip is flipped circuit-side down and connected via bumps. Wires can only come from the chip's edges, but bumps can be placed by the thousands across the whole surface, greatly increasing terminal count. Connections are also shorter, which helps high-speed signals. The gap between chip and substrate is filled with underfill resin to protect the bumps from cracking.

Unlike the front-end, the back-end is essentially one-way with no repeated loop, though electrical testing comes first. After WAT, wafer testing applies fine needles to each chip still on the wafer to check whether it works; defective chips are excluded from assembly. Then come back-grinding, dicing (the point where handling shifts from whole wafers to individual chips), die bonding onto a substrate or lead frame, electrical connection by wire bonding or flip-chip, molding in resin, finishing such as marking, and the final test. Ichiken notes that for cutting-edge chips this standard flow is starting to change, which they return to later.

Wafer Testing: Testers, Probe Cards, and Probers

Wafer testing exists because assembling defective chips wastes effort and materials downstream. A probe card, a bundle of fine needles, touches each chip's terminals, and a tester sends signals and reads the responses. The tester and wafer are connected through the probe card. Since a wafer holds hundreds of chips, a prober moves the wafer precisely to position each one. Well-known tester makers are Advantest and Teradyne; prober makers include Tokyo Electron and Tokyo Seimitsu. The two machines are used together as a unit, and the tester is used again in the final test.

Ichiken describes the tester as essentially a collection of measuring instruments: power supplies that apply voltage and measure current, circuits that send signals to each terminal and evaluate the response, and circuits that generate timing at picosecond resolution, replicated for hundreds to thousands of terminals and operated simultaneously. Because testing takes time, the key is how fast it can be finished. Components include probe cards with hundreds of pins and measurement circuits.

47:40

Back-End Steps: Grinding, Dicing, Bonding, Connecting, Molding

Back-grinding thins the wafer, which was kept thick through the front-end to prevent breakage but won't fit thin packages as is, especially for stacked chips such as HBM. Protective tape covers the circuit side, and a rotating grinding wheel thins the back. Disco leads the world in equipment and also makes its own grinding wheels. Ichiken explains three roles of back-grind tape. It protects the circuit from grinding debris and the cooling water. It absorbs surface unevenness and holds the wafer flat; since any variation in tape thickness becomes variation in wafer thickness, the tape needs high precision. And it must peel off cleanly: because thinned wafers break easily, a tape whose adhesive weakens under UV light is used so it comes off without residue or force. Lintec and Mitsui Chemicals make such tapes.

Dicing cuts hundreds to thousands of chips apart while still on tape, commonly with a high-speed diamond blade and water spray. Disco and Tokyo Seimitsu are said to account for most blade dicing equipment. Dicing tape, like back-grind tape, comes from Japanese companies such as Lintec.

Die bonding picks chips one by one off the tape and fixes them to a package substrate or lead frame, leaving chips marked defective in wafer testing behind. Besi in the Netherlands and Singaporean companies are strong in equipment. For lead frames and adhesive films, Ichiken again remarks on Japanese strength in materials. High-heat chips such as power semiconductors may use solder or silver sintering for better heat dissipation instead of adhesive.

Electrical connection uses either wire bonding, attaching gold or copper wires thinner than a hair one at a time, or flip-chip, which connects bumps directly and suits CPUs and GPUs because the short wiring carries signals quickly. Equipment includes wire bonders and flip-chip bonders. Tanaka is a bonding wire supplier, and underfill for the gaps between bumps comes from companies such as Namics and Resonac.

Molding encases the chip in resin because exposed chips and wires are vulnerable to moisture, dust, impact, and light. The substrate is clamped in a mold and resin is injected and cured, using methods such as compression molding and transfer molding. TOWA has a large share and is strong in compression molding, where resin is pressed down from above, a method considered suited to encapsulating thin stacked chips such as HBM. The encapsulant is epoxy resin with fillers; Sumitomo Bakelite is considered the world leader, followed by Resonac and Shin-Etsu Chemical. Ichiken again notes the strength of Japanese materials makers.

Finishing includes laser marking and terminal preparation; BGA packages, with solder balls on the back, allow very dense layouts. In the final test, packaged chips are placed in sockets and driven by the tester. Test sockets are a component category, and some companies take on testing as contract work. Companies that handle the whole back-end as a package are OSATs, and ASE is the industry leader. With AI chips integrating multiple dies, though, advanced packaging has become so important that front-end companies such as TSMC and Intel have started doing packaging themselves.

53:48

Why the Back-End Is Now in the Spotlight

Ichiken explains that chip performance has traditionally improved in two ways: miniaturization, which packs more transistors into the same area, and making chips larger. Both have hit walls. Each generation of miniaturization requires larger equipment and processes, and costs are rising rapidly. Larger chips are more likely to be ruined by a single defect, so yield falls.

Hence chiplets: build small dies by function, select only the good ones, and reconnect them at high density side by side or stacked inside the package. This makes the chip effectively larger. In AI GPUs, HBM is placed right next to the GPU so large amounts of data move over short distances.

Ichiken then walks through the package substrate. The center core is glass cloth impregnated with resin, which resists heat-induced warping; Nittobo's T-glass is commonly used. Above and below are build-up layers of alternating insulating film and copper wiring; the insulating film is Ajinomoto's ABF, and the wiring uses Mitsui Mining & Smelting's ultra-thin copper foil. Spacing widens layer by layer to connect the chip's fine terminals to the large solder balls below. The outermost green layer is solder resist, protecting the wiring. Ichiken says that the larger the AI chip, the more substrate materials and build-up quality matter. Ibiden is one manufacturer of these substrates. Ichiken acknowledges this section was rushed and incomplete and plans a separate, thorough video on packaging.

On how chips are connected inside the package, Ichiken says AI chip performance depends not just on a single chip but on how closely multiple chips can be connected and how many wiring paths can be added. Flip-chip is the standard for high-performance chips. Many AI GPUs go further: GPU and HBM sit side by side on a silicon interposer. This is needed because the green organic substrate can't achieve silicon's precision in wiring fineness and via spacing, which makes connecting GPU and HBM directly on it difficult, especially given the large number of connections. Precision-machined silicon links GPU and memory, and the interposer then connects to the organic substrate and ultimately to the PCB. Because the back-end now helps determine the chip's performance itself, Ichiken concludes, more sophisticated technologies closer to front-end processing are moving into it.

57:54

PCB Assembly and Why MLCCs Matter for AI

Finished chips only fulfill their function once mounted with other components on product boards. Mounting machines (placement machines) are an area where Japanese makers are strong: FUJI, Yamaha Motor, Panasonic Connect, and JUKI. Japanese companies are also competitive in electronic components around the chip.

Ichiken singles out MLCCs (multilayer ceramic capacitors), thousands of which sit on AI GPU boards. When an AI GPU starts computing, billions of transistors switch at once, and currents of several hundred amperes flow over nanosecond timescales, causing voltage drops in the wiring to the chip. Meanwhile the chip runs at around 1V and tolerates fluctuations of only a few tens of millivolts; beyond that it won't work properly. The power supply responds only in milliseconds, so MLCCs are placed near the GPU to deliver charge instantly, suppressing voltage fluctuation and keeping power supply impedance low. Ichiken considers this important enough to cover in another video.

Conclusion

Ichiken's takeaway is that making a single chip is divided among many countries and companies, each handling its own part of the process, and that changing even one step can affect the steps that follow. They present the slides as a summary of what they learned through their research and invite viewers to add further information in the comments for others' benefit. Package substrates and MLCCs remain open topics they intend to cover in future videos.