# Intel Gaussian and Neural Accelerator
**Entity class:** Low-power neural coprocessor
## Definition and history
Intel's Gaussian and Neural Accelerator, or GNA, is a low-power device for continuous inference workloads such as speech processing and noise reduction. [[articles/Intelligence Inside|Intelligence Inside]] locates it inside shipping processors while the principal CPU can remain in a lower-power state. A neural model running on that coprocessor performs inference on live input without requiring a user-facing chatbot. The chip's deployment illustrates the distinction between an always-available inference layer and the public chronology of conversational AI.
## Mechanism and limits
The accelerator executes a model whose training normally occurs elsewhere; its local operation is inference rather than proof of on-device weight updates. Offloading matters because continuous sensing would impose unacceptable energy or wake-up costs on a main processor. Specific generations and supported workloads vary, so the device family should not be reduced to a single fixed capability.
## Relationships
**Manufacturer:** [[wiki/Intel|Intel]]. **Function:** [[wiki/Inference Layer|Inference Layer]]. **Comparison:** [[wiki/Always On Processor|Always On Processor]] and [[wiki/Intel Thread Director|Intel Thread Director]]. **Article route:** [[articles/Intelligence Inside|Intelligence Inside]].
## Sources
[Intel OpenVINO GNA device documentation](https://docs.openvino.ai/2023.3/openvino_docs_OV_UG_supported_plugins_GNA.html); [[articles/Intelligence Inside|Intelligence Inside]].