ASIC vs FPGA: what's the difference?

An ASIC, short for application-specific integrated circuit, is a chip whose circuits are permanently manufactured for one task, and it will never do anything else. An FPGA, a field-programmable gate array, is a chip covered in standard circuit blocks whose connections are set by loading a file, so the same chip can become one circuit today and a different one tomorrow. The difference between them is commitment. Committing the design makes each chip cheaper, smaller and more efficient once volumes are large; staying changeable costs more per chip, but a design error is corrected by loading a new file. So the choice is mostly about quantity, and the crossover is often put somewhere between 5,000 and 50,000 units.

An ASIC, short for application-specific integrated circuit, is a chip manufactured so that its circuits perform exactly one task. An FPGA, a field-programmable gate array, is a ready-made chip whose internal connections are set afterwards, by loading a configuration file. Both begin the same way, with engineers describing the wanted circuit in a design language. For an ASIC, that design is sent to a factory and manufactured permanently into silicon, which commonly costs millions of dollars before the first chip exists. For an FPGA, the design is loaded onto a chip bought off the shelf. After that the economics reverse. Each manufactured ASIC costs a fraction of an FPGA and does its task with less power, because the blocks and wiring that keep an FPGA changeable take up silicon, add delay and draw power. A flawed FPGA design is corrected by loading a fixed file, often the same day; a flawed ASIC means a new manufacturing run, months away, with the faulty chips already built. So the choice comes down to volume and certainty: past a crossover often put between 5,000 and 50,000 units, with a design that is trusted, the ASIC repays its upfront cost. Neither is a processor. A GPU's circuits are fixed but run changeable software, whereas reconfiguring an FPGA changes the circuit itself, and an ASIC's circuit never changes at all. Google's TPU is a famous ASIC, built only for neural-network arithmetic, but most are anonymous: the fixed chips inside appliances, cars and phones, made in the millions for one job each.

An ASIC is like a factory mould; an FPGA is like a box of snap-together parts. Machining the mould costs a fortune, takes months, and produces exactly one shape. After that, every part it presses out is identical, exactly right, and costs a tiny fraction of what the mould did. If the shape turns out to be wrong, another mould has to be machined. The kit works the other way around. Its parts cost more and the result is bulkier and weaker, but it comes apart tonight and rebuilds as something else, so a mistake costs an evening, not another mould. The mould is the ASIC, committed and cheap at volume; the kit is the FPGA, changeable and dear per part; and the choice depends on how many you need and how sure you are.