Neural Engine
An AI inference framework developed by Blackmagic Design for DaVinci Resolve, optimized for video processing workflows. It supports multiple compute platforms including NVIDIA CUDA, AMD ROCm, and Apple Silicon, with deep adaptation for color science and video codec operations.
Core Facts
Since the A11 Bionic chip, Apple has built in a Neural Engine to accelerate machine learning inference tasks.
The M4 chip's Neural Engine can perform 38 trillion operations per second (38 TOPS)
Apple Intelligence relies on Apple's self-developed Neural Engine to run a small language model of approximately 3 billion parameters on-device for local tasks, while accessing larger-scale models via a private cloud computing architecture in the cloud.
Timeline (last 90 days)
Early Neural Engines could only be invoked indirectly through Core ML, with a limited set of expressible operators; only newer platforms gradually opened up lower-level access.
The cost of the neural engine is a more restricted programming model, with operator support and memory layout both requiring adaptation to the hardware.
In continuous inference scenarios, the neural engine is more power-efficient and generates less heat than the GPU, making it suitable for long-duration conversational applications.
Typical representatives of edge AI include Google's TensorFlow Lite, Apple's Neural Engine and Qualcomm's AI Engine.
The neural engine integrated into M-series chips is specially optimized for matrix operations and low-precision inference, and can accelerate model computation at low power consumption.
The neural engine is an independent computing unit designed specifically for machine learning tasks, skilled at operations such as matrix multiplication and convolution, and its power consumption is far lower than general-purpose computing on CPU or GPU.
One of the advantages of Apple's Neural Engine lies in its deep integration with the Core ML framework, enabling developers to more efficiently offload model inference to dedicated hardware.
At a 140k context, invoking the phone's Neural Engine reduced the write time per token from 279 ms to 176 ms.
The Neural Accelerator is a dedicated hardware unit first introduced by Apple in the M5 series chips, designed specifically to accelerate neural network inference tasks, operating independently of the CPU and GPU, and continued and integrated into the next-generation GPU in the M6.
Macs equipped with Apple Silicon (M-series chips) have a dedicated Neural Engine that can run language models with tens of billions of parameters at low power consumption.
17 more timeline events
All Facts (20)
Since the A11 Bionic chip, Apple has built in a Neural Engine to accelerate machine learning inference tasks.
90%VerifiedThe M4 chip's Neural Engine can perform 38 trillion operations per second (38 TOPS)
90%VerifiedApple Intelligence relies on Apple's self-developed Neural Engine to run a small language model of approximately 3 billion parameters on-device for local tasks, while accessing larger-scale models via a private cloud computing architecture in the cloud.
75%VerifiedApple's Neural Engine first appeared with the A11 Bionic chip in 2017, with the A11 capable of executing 600 billion operations per second, while the A18 Pro chip has reached 35 trillion operations per second.
75%VerifiedThe Neural Accelerator is a dedicated hardware unit first introduced by Apple in the M5 series chips, designed specifically to accelerate neural network inference tasks, operating independently of the CPU and GPU, and continued and integrated into the next-generation GPU in the M6.
65%VerifiedApple's self-developed chips have a built-in neural engine, and the latest M4 series delivers up to 38 TOPS of AI inference performance.
65%VerifiedThe M4 chip's Neural Engine delivers a peak compute performance of 38 TOPS (38 trillion operations per second)
65%VerifiedThe Neural Engine in Apple's custom chips is an NPU optimized specifically for matrix multiplication and convolution operations, and offers higher energy efficiency than general-purpose CPUs and GPUs on specific AI inference tasks.
65%Verified苹果M系列芯片中的Neural Engine本质上属于NPU
65%Unverified苹果从A11仿生芯片开始持续提升端侧AI算力
90%UnverifiedApple Silicon Neural Engine的峰值算力从M1的11 TOPS跃升至M4的38 TOPS，性能提升超过3倍
90%UnverifiedFoundation Models框架将大语言模型推理能力下沉到设备本地执行，而非依赖云端API调用
90%UnverifiedThe Neural Engine integrated in the M2 Ultra contains 32 dedicated cores with a theoretical compute performance of up to 31.6 TOPS.
60%UnverifiedIntel's Meteor Lake and subsequent architectures, AMD's Ryzen AI series, and Apple's Neural Engine all integrate NPU modules.
60%UnverifiedNeural Engine的设计哲学源自谷歌TPU（张量处理单元）的架构思路，但针对移动端和桌面端的功耗约束进行了优化
60%UnverifiedEarly Neural Engines could only be invoked indirectly through Core ML, with a limited set of expressible operators; only newer platforms gradually opened up lower-level access.
50%UnverifiedIn continuous inference scenarios, the neural engine is more power-efficient and generates less heat than the GPU, making it suitable for long-duration conversational applications.
50%UnverifiedThe cost of the neural engine is a more restricted programming model, with operator support and memory layout both requiring adaptation to the hardware.
50%UnverifiedTypical representatives of edge AI include Google's TensorFlow Lite, Apple's Neural Engine and Qualcomm's AI Engine.
50%UnverifiedThe neural engine integrated into M-series chips is specially optimized for matrix operations and low-precision inference, and can accelerate model computation at low power consumption.
50%