AppsGuidesAI DictionarySmall FixesOn RecordAboutFollow the lab

FPGA

Also called field programmable gate array

An FPGA is a chip whose wiring can be rearranged after it is made, so engineers can turn it into almost any custom circuit, including an AI engine.

Chips and scaleTechnical

In plain words

Most chips are fixed forever when they leave the factory. A GPU or a phone processor does what it was designed to do, and nothing else. An FPGA, short for field-programmable gate array, is different: it is a grid of blank building blocks whose connections are set later, with software.

Load one design and the FPGA becomes a video filter. Load another and it becomes a network switch, or a small engine for running an AI model. The same chip can be rewired again and again, even after it is installed out in the field, which is where the name comes from.

For AI, that flexibility helps when answers must come back very quickly, when a design needs a custom number format, or when it may change. The price is efficiency: a rewirable chip is bigger, slower and hungrier than a fixed chip built for the same job. So most large-scale AI runs on GPUs and other AI accelerators instead.

Everyday example

LEGO bricks versus a moulded toy

A toy car from a factory mould is cheap, strong and only ever a car. A box of LEGO can be a car today and a crane tomorrow, but the LEGO car is chunkier and wobblier than the moulded one. An FPGA is the LEGO box of chips: one set of blocks, rebuilt for each job. A fixed chip designed for one task, called an ASIC, is the mould.

Where the example breaks: LEGO is rebuilt by hand. An FPGA is rewired by loading a file, called a bitstream: nothing physically moves, only tiny electronic switches change setting.

Go deeper

How it actually works

Inside, an FPGA is a sea of small logic blocks surrounded by wiring that can be switched on or off.

  1. Lookup tables do the logic

    Each logic block holds small lookup tables: memory cells that can be set to act as any small logic function, plus a flip-flop that stores one bit. A grid of these blocks covers the chip, joined by programmable routing channels.

  2. Hard blocks for the maths

    Modern FPGAs also include fixed blocks for multiplying numbers (DSP blocks) and on-chip memory. AI designs lean on these, because neural networks spend most of their time multiplying and adding numbers.

  3. Describing the circuit

    Engineers write the design in a hardware description language such as Verilog or VHDL. Software tools place it onto the blocks and produce a bitstream, a file that sets every switch. Most FPGAs load it each time they power up.

  4. Why choose one for AI

    FPGAs can answer without waiting to fill a batch, and can use unusual number sizes. Microsoft's Project Brainwave (2017) ran a large neural network on Intel FPGAs in under a millisecond per request, and Microsoft had used FPGAs to speed up Bing search since 2015.

  5. The cost of flexibility

    A classic 2007 study compared FPGAs with fixed chips (ASICs) doing the same work. The FPGAs needed about 35 times the silicon area, were three to four times slower and used about 14 times the dynamic power. Built-in hard blocks narrowed the area gap to as little as 18 times.

Invented
1984, by Ross Freeman, who co-founded the chip company Xilinx
First product
Xilinx XC2064, 1985, with 64 programmable logic functions
Main makers
AMD, which completed its purchase of Xilinx in February 2022, and Altera, 51% owned by Silver Lake since 2025 with Intel keeping 49%
Uses
Software-defined radios, data center routers, neural networks and even recreating old computers (IEEE Spectrum)

Questions people ask

What does FPGA stand for?

Field-programmable gate array. Field-programmable means it can be programmed after it ships, out in the field. Gate array refers to the grid of logic blocks inside, which together act like the logic gates of an ordinary circuit.

Is an FPGA better than a GPU for AI?

For most AI work, no. GPUs and fixed AI chips are faster and use less power for big models. FPGAs win in narrower cases: very low delays without batching, custom number formats, or designs that must change after the hardware is installed.

What is the difference between an FPGA and an ASIC?

An ASIC is a chip designed for one task and fixed at the factory. An FPGA can be reprogrammed as often as you like. ASICs are smaller, faster and more efficient, but designing one costs far more and takes much longer.

Can an FPGA be reprogrammed after it is sold?

Yes, that is the point of it. Hardware updates can be sent like software updates: load a new bitstream and the chip becomes a different circuit. This lets companies fix bugs or add features to devices already in use.

Was this clear?

1 is confusing, 5 is crystal clear. One tap, no account needed.

Spotted something wrong, or know a better example? Suggest a change

Further reading

Prepared by Mundhra Labs. Reviewed by Aditya Mundhra. Checked November 2026.