ASIC
In plain wordsA 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.
An integrated circuit whose logic is fixed when it is made and tailored to one application or family of workloads, in contrast to a CPUCPU, abbr.: The main general-purpose chip in a computer, good at running many different programs step by step. In AI systems it usually coordinates the work while specialized chips do the heavy math., which can run any program, or an FPGAFPGA, abbr.: A 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., which can be rewired after fabrication. Tailoring the datapath, memory hierarchy and number formats to one workload yields better performance per watt and per dollar, at the cost of high non-recurring engineering expense (which can exceed $100 million at leading nodesprocess node, n.: The label for a generation of chip-making technology, like '3 nanometer'. Smaller numbers mean newer, denser chips, but the number no longer measures any actual part of the chip.), multi-year design cycles and the risk that workloads change before the chip ships. In AI the term usually means custom accelerators designed by or for large users — Google's TPUTPU, abbr.: Google's own family of chips built specifically for AI, used to train and run its Gemini models and rented to other companies through its cloud., Amazon's Trainium and Inferentia, Microsoft's Maia and Meta's MTIA — often co-developed with design-services firms such as Broadcom or Marvell and manufactured by a semiconductor foundrysemiconductor foundry, n.: A factory company that manufactures chips designed by other companies. Taiwan's TSMC, the largest, makes nearly all of the world's most advanced AI chips.. Bitcoin-mining chips are another familiar class.
Origin
The ASIC industry grew from the gate arrays of the late 1960s and 1970s and from standard-cell design, which let customers specialize partly prefabricated or library-based chips; the term was common by the 1980s. Custom AI ASICs became a major industry trend after Google revealed its TPU in 2016.
Usage note
Industry usage often contrasts custom ASICs with merchant 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., although a modern GPU is itself an application-specific design. A TPUTPU, abbr.: Google's own family of chips built specifically for AI, used to train and run its Gemini models and rented to other companies through its cloud. is an ASIC; an FPGA is not.
Related
- 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.
- TPUTPU, abbr.: Google's own family of chips built specifically for AI, used to train and run its Gemini models and rented to other companies through its cloud.
- FPGAFPGA, abbr.: A 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.
- 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.
- semiconductor foundrysemiconductor foundry, n.: A factory company that manufactures chips designed by other companies. Taiwan's TSMC, the largest, makes nearly all of the world's most advanced AI chips.
- NPUNPU, abbr.: A small AI chip, or part of a chip, built into phones, laptops and other devices so they can run AI features quickly without draining the battery or using the cloud.
- FPGAFPGA, abbr.: A 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.
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., 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., NPUNPU, abbr.: A small AI chip, or part of a chip, built into phones, laptops and other devices so they can run AI features quickly without draining the battery or using the cloud., semiconductor foundrysemiconductor foundry, n.: A factory company that manufactures chips designed by other companies. Taiwan's TSMC, the largest, makes nearly all of the world's most advanced AI chips., TPUTPU, abbr.: Google's own family of chips built specifically for AI, used to train and run its Gemini models and rented to other companies through its cloud., FPGAFPGA, abbr.: A 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.