Also known as: ASIC

An application-specific integrated circuit (ASIC) is an integrated circuit custom-designed for a single, fixed task — trading away general-purpose flexibility for the best achievable speed, power efficiency, and per-unit cost.¹

Overview

Where a CPU or FPGA is built to run many possible workloads, an ASIC’s circuit is etched permanently into silicon to do exactly one job — a network switch chip, a Bitcoin miner, a phone’s modem, a TPU. The design is fabricated at a foundry such as TSMC from a set of photomasks; this non-recurring engineering cost is large and a finished ASIC cannot be changed, so the economics only work at high volume or where nothing else meets the performance target.

Trade-offs

The classic spectrum runs CPU → GPU → FPGA → ASIC, with flexibility falling and efficiency rising at each step. An ASIC wins decisively on performance-per-watt and cost-at-scale but is inflexible and slow to bring to market; an FPGA is the hedge when volumes are low or the design may still change. Many specialized hardware accelerators are ASICs. For a project like GopherTrunk, ASICs already appear inside the radio — the RTL-SDR’s tuner and demodulator are fixed-function chips — even though the decoding itself is done in software.

Sources

1. Application-specific integrated circuit — Wikipedia, on custom fixed-function chips and their trade-offs. ↩

See also

• Field-programmable gate array (FPGA)
• Integrated circuit
• Tensor processing unit (TPU)
• Hardware acceleration
• SoC vs discrete
• Semiconductor