An artificial intelligence accelerator is a processor or processing unit optimized for AI workloads. It provides high parallel throughput, fast memory movement and numerical formats suited to neural-network operations that would be slower or less efficient on a general-purpose CPU.
Accelerators include GPUs and dedicated neural-processing chips. Model performance depends on software support as well as hardware, because kernels, compilers, memory layout and serving frameworks must use the device effectively. Supporting more accelerator vendors can reduce dependence on one hardware ecosystem.
Acronyms and aliases
AI chip synonymAI accelerator variant
Specialised terms
Related terms
Frequently asked questions
What kinds of hardware are AI accelerators?
GPUs, tensor-processing units, neural-processing units and other application-specific chips can serve as AI accelerators. Their strengths differ across training, inference and model architectures.
Why do AI models need accelerators?
Neural networks perform many parallel matrix operations. Accelerators execute those operations with higher throughput and often better energy efficiency than general-purpose processors.
