Field-Programmable Gate Array (FPGA)

A field-programmable gate array (FPGA) is an integrated circuit that can be configured by its user after it leaves the factory. Instead of having fixed wiring for one job, an FPGA contains a large grid of programmable logic blocks and connections that engineers can arrange to behave like a custom chip. The "field-programmable" part of the name means it can be set up, and later changed, in the field rather than only at the point of manufacture.

FPGAs sit between general-purpose processors and fully custom chips. A central processing unit or graphics processing unit runs software on fixed hardware, while an application-specific integrated circuit (ASIC) is designed for one task and cannot be changed. FPGAs offer much of an ASIC's speed for a particular job while keeping the flexibility to be reprogrammed. They are used in telecommunications equipment, aerospace, data centres, medical imaging, financial trading systems and, for a short period, cryptocurrency mining.

A normal computer chip is like a kitchen with fixed appliances: it can cook many dishes, but every recipe has to work with the same oven and hob. A custom chip is like a factory built to make one product very quickly. An FPGA is like a kitchen full of modular units that can be rearranged overnight into a production line for whatever dish is needed next.

That flexibility makes FPGAs useful when a task must be done extremely fast but the exact method may change, or when building a custom chip would be too costly. The price is that setting one up needs specialist skills, and for very large production runs a custom chip usually works out cheaper and more efficient.

An FPGA contains thousands to millions of small logic elements, each able to perform simple operations, along with memory blocks, arithmetic units and input and output connections. Engineers describe the circuit they want in a hardware description language such as Verilog or VHDL. Design software then converts that description into a configuration file, often called a bitstream, which sets how the logic elements are linked. Loading a new bitstream changes what the chip does.

Because the circuit runs in hardware, many operations happen at the same time rather than one after another, which gives FPGAs high speed and predictable timing for suitable tasks. The best-known makers include Xilinx, which became part of AMD in 2022, and Altera, which was bought by Intel in 2015. Smaller vendors serve niche and low-power markets.

In the early years of Bitcoin, which secures its network through proof of work, miners moved from ordinary processors to graphics cards and then, around 2011 and 2012, to FPGA boards, which could compute the network's hash function more efficiently per unit of electricity. The FPGA phase was short. From 2013, ASIC miners built only for Bitcoin's SHA-256 algorithm arrived and were far faster and more efficient, pushing FPGAs and GPUs out of Bitcoin mining as mining difficulty climbed.

FPGAs have continued to appear in mining for networks whose algorithms were designed to resist ASICs, or that changed their algorithms to shut out existing ASIC hardware, as described under ASIC resistant. Because an FPGA can be reprogrammed, a miner can switch to a different coin's algorithm when profitability changes, something a fixed mining rig built on ASICs cannot do. Some hardware projects in zero-knowledge proof generation also use FPGAs to speed up heavy cryptographic calculations.

Trading firms use FPGAs where every microsecond matters. In high-frequency trading, an FPGA placed on a network card can read market data, apply simple trading logic and send an order faster than software running on a general-purpose server. Exchanges and market data vendors also use FPGAs to handle and filter large volumes of messages, and banks use them to accelerate risk and pricing calculations. The speed advantage is part of a wider technology race in algorithmic trading.

• Speed with flexibility. FPGAs can perform specific tasks far faster than general processors while remaining reprogrammable.
• Lower upfront cost than custom chips. Designing an ASIC can cost millions, while an FPGA can be bought off the shelf and configured.
• Specialist skills required. Hardware design is slower and harder than writing software, and errors can be costly to find.
• Less efficient at scale. For a stable task produced in large numbers, an ASIC usually beats an FPGA on speed, power use and cost per unit.

These trade-offs explain why FPGAs often serve as a bridge: used while a design is still changing or a market is too small for custom chips, then replaced by ASICs once a task becomes fixed and large enough.