Fabless. The transformation of the semiconductor industry (D. Nenni, P. McLellan)

Explain the fabless semiconductor ecosystem. What it consists of and how it works.

• IC production:
  • One wafer contains many chips (depending on their size).
  • IC are deposited on the wafer in layers, for all the chips at the same time.
  • The equipment is universal: it doesn't need to be changed for a different design to be printed.
• ASICs
  • Systems companies wanted custom components, not the generic stuff available in the market, so ASIC companies filled this niche or producing chips based on a functional spec.
  • The process looks like this:
    • Systems company picks an ASIC supplier and receives their design tools and the library of basic building blocks.
    • They build the functionality they want out of them, using simulation for testing the design.
    • ASIC company produces test samples, they are installed onto a board and the system is tested.
    • If everything works, it goes to mass manufacturing and gets shipped.
  • The leaders were VLSI Technology and ISI Logic. Both were later acquired.
  • What remains of ASIC is Design Services companies that use foundries for manufacturing.
• FPGA
  • Kind of like programmable ASICs. More expensive but no fixed costs. Also much faster.
  • Because of their regularity and uniformity, FPGAs are useful for the foundries to test and improve new proces nodes at early stages.
  • Software for programming FPGAs is mostly provided by FPGA vendors. EDA tried it, but the economics are worse than with software for normal chip design.
  • There's an FPGA-specific IP industry.
  • FPGA is dominated by Xilinx and Altera and entering the market has been hard. One of the reasons is patents.
  • FPGA market is currently mostly driven by Adaptive Compute Acceleration Platforms for edge AI applications.
  • FPGAs can also be supplied as IP (to be embedded in a bigger SoC) or as chiplets (to be integrated with a SoC that occupies the same enclosure).
• Move to the fabless model
  • The cost of fabs is going up with more advanced process nodes, so only very few companies can afford their own fabs. Selling fab services created nice economies of scale and enabled creation of pure-play foundries (TSMC was the first in 1978).
  • As outsourcing of manufacturing grew, fabless companies became possible. The relative simplicity and speed of starting fabless ensured that many startups chose this route. Eventually most of the business of TSMC would come from fabless companies.
  • As the costs and complexity of building modern fabs grew, most companies were forced to outsource leading process nodes to foundries. Later even most foundries couldn't keep up and in 2020 the top process nodes are dominated by TSMC with some competition from Samsung. Everyone else is basically out.
• EDA
  • Originally photolitography masks used for producing chips were made by hand (and then photoreduced). However, as density grew and chip sizes (in transistors) increased, this became unwieldy. Automation was necessary.
  • First EDA tools were hardware, but soon software followed. Then the tools became more and more advanced and functional and the industry consolidated.
  • The industry is now dominated by Cadence, Synopsis and Mentor Graphics, but new companies with new technologies appeared regularly -- they just get bought by the majors when their tech is proven in the market. Lately more innovation is happening inside of the big companies because the toolset is bigger and more complicated, making it harder to enter the market.
• IP
  • As scale of integration grew and more and more components were fit on the same chip, it became economical to only design the distinguising features of SoC inhouse and buy existing designs for other components. Thuse the IP (intellectual property) industry started and it kept growing since.
  • One very successful IP company is ARM, whose designs power almost all modern smartphones.
• What's next?
  • New processes (non-silicon) might create opportunities for new entrants into the foundry business. However, we have not run out of silicone potential (peak growth is predicted for 2038-ish).
  • Demand for chips is strong. With latest development in AI, IoS, and automotive applications, on top of the more traditional computers and smartphones, it seems that it will continue.
  • China will probably play a bigger role.