FPGA
In plain wordsA chip whose circuits can be rewired by software after it is made, so engineers can turn it into a custom processor — slower than a purpose-built chip, but flexible.
An integrated circuit whose logic and wiring can be reconfigured after manufacture by loading a configuration bitstream: a grid of lookup tables, flip-flops, hard multiply–accumulate (DSP) blocks and on-chip RAM, linked by programmable routing. Designers describe circuits in hardware description languages or high-level synthesis tools and can build custom datapaths with arbitrary bit widths. In AI, FPGAs suit low-latency inference, network processing and prototyping of ASICASIC, abbr.: A chip whose circuits are laid out to do one kind of work extremely well instead of running any program. AI companies design their own to run their models more cheaply than with off-the-shelf chips. designs; Microsoft's Catapult and Brainwave projects used them to accelerate Bing search and to serve neural networks in its data centers. Because they trail GPUsGPU, abbr.: A computer chip first designed to draw video-game graphics that turned out to be ideal for AI, because it can do thousands of simple calculations at the same time. and ASICs in peak throughput and are harder to program, FPGAs hold a niche rather than a large share of AI compute.
Origin
Xilinx shipped the first commercially viable FPGA, the XC2064, in 1985, following programmable-logic devices such as Altera's EP300 (1984). Intel bought Altera in 2015 and AMD acquired Xilinx in 2022.
Related
- ASICASIC, abbr.: A chip whose circuits are laid out to do one kind of work extremely well instead of running any program. AI companies design their own to run their models more cheaply than with off-the-shelf chips.
- AI acceleratorAI accelerator, n.: Any chip built or tuned to speed up AI — graphics chips, Google's TPUs, phone 'neural engines' and many startup designs. Such chips do the heavy math of training and running models.
- GPUGPU, abbr.: A computer chip first designed to draw video-game graphics that turned out to be ideal for AI, because it can do thousands of simple calculations at the same time.
- systolic arraysystolic array, n.: A grid of tiny calculators that pass numbers to their neighbors in rhythm, like a heartbeat, so each value is reused many times. AI chips use it to multiply large tables of numbers efficiently.
- ASICASIC, abbr.: A chip whose circuits are laid out to do one kind of work extremely well instead of running any program. AI companies design their own to run their models more cheaply than with off-the-shelf chips.
Sources
- Putnam, A. et al. (2014). A Reconfigurable Fabric for Accelerating Large-Scale Datacenter Services. Proceedings of ISCA 2014.
Mentioned in
AI acceleratorAI accelerator, n.: Any chip built or tuned to speed up AI — graphics chips, Google's TPUs, phone 'neural engines' and many startup designs. Such chips do the heavy math of training and running models., ASICASIC, abbr.: A chip whose circuits are laid out to do one kind of work extremely well instead of running any program. AI companies design their own to run their models more cheaply than with off-the-shelf chips.