Skip to main content

Technology

ASICs

Application-Specific Integrated Circuits custom-designed for dedicated tasks like crypto mining, AI inference, and networking.

199 companies

Explainer

ASICs explained

An ASIC, or application-specific integrated circuit, is a chip designed to do one job or a narrow set of jobs. Its circuits are fixed when the chip is made. A CPU can run any program; an ASIC does what it was designed for and nothing else.

Designing an ASIC means writing the logic in a hardware description language, verifying it, and then paying a foundry to manufacture it. The first run has a large one-time cost, called non-recurring engineering (NRE), which is paid back only if enough chips are sold or deployed.

How it differs from the usual approach

A CPU or GPU spends silicon and energy on flexibility. An ASIC removes what the job does not need, so it can be much faster or more energy efficient at that job. The trade is cost and risk. The design cannot be changed after manufacture, and if the target algorithm changes, the chip loses value. An FPGA sits in between, since it can be reprogrammed but is less efficient.

Where it is used

  • Bitcoin mining machines, which compute one hash function.
  • AI accelerators designed by cloud providers for their own workloads.
  • Network switches and base station chips.
  • Video encoding and decoding in phones and data centers.
  • Automotive, medical and consumer devices produced in large volumes.

What to watch

  • Cloud providers work with design partners such as Broadcom and Marvell to build custom AI ASICs.
  • Rising manufacturing and design costs at advanced process nodes favour large buyers who can spread the cost over many chips.
  • AI models change quickly, which is the main risk for fixed-function AI chips.
  • Chiplet-based designs lower the cost of making new ASICs by reusing proven dies.

Frequently asked questions

What is the difference between an ASIC and an FPGA?

An ASIC has fixed circuits set at manufacture and is the most efficient at its task. An FPGA can be reprogrammed after it is made, which gives flexibility at a cost in speed, power and unit price.

Why are ASICs expensive to start?

Design, verification and the masks used in manufacturing carry a large one-time cost. That cost only makes sense if the chip is produced in volume or has a clear performance benefit.

Are AI chips ASICs?

Many are, in a broad sense. Google's TPU and Amazon's Trainium are custom chips for AI. GPUs are more general and are not usually called ASICs.