An FPGA is a grid of small logic blocks and programmable wiring. Engineers describe the circuit they want in a hardware design language, and a tool translates that into a configuration file that sets up the chip. Load a different file and the same chip becomes a different circuit. That flexibility makes FPGAs popular for prototyping chips, in telecom equipment and in finance, where traders want answers in microseconds.
For AI, FPGAs sit between the two better-known options. A GPU is flexible and easy to program through frameworks but draws a lot of power. An ASIC is the most efficient for one job but takes years and a large upfront investment to design, and cannot be changed afterward. An FPGA can be tuned to one model for fast, predictable latency and then reprogrammed when the model changes. Microsoft’s Project Brainwave is a well-documented case: FPGAs in its data centers served real-time deep learning inference.
The downside is effort. Programming an FPGA well needs hardware-design skills most AI teams lack, and the software ecosystem is far thinner than CUDA’s. For training today’s large models, GPUs and AI ASICs dominate. FPGAs show up more in specialized inference, networking and edge AI devices.
FPGAs have not been discussed on Chain of Thought so far; this entry relies on the sources listed.
Sources
- Microsoft Research: Project Catapult — Microsoft's project combining programmable hardware and software, using field-programmable gate arrays (FPGAs) in its data centers to deliver performance improvements; checked October 24, 2026.
- Microsoft Research: Project Brainwave — Describes a soft neural processing unit built on a high-performance FPGA that accelerates deep neural network inference for real-time AI in the cloud and at the edge; checked October 24, 2026.
- Lacey, Taylor and Areibi: Deep Learning on FPGAs: Past, Present, and Future (2016) — Survey of using FPGAs to accelerate deep learning, discussing their reconfigurability and performance per watt as an alternative to GPUs; checked October 24, 2026.