（認(rèn)知有限，如有不妥，還望告知，歡迎討論）

先寫總結(jié)

（從ML&DL角度）

CPU： 主要用來做串行計(jì)算的，可以處理相對(duì)復(fù)雜的運(yùn)算

GPU：主要用來做并行計(jì)算和處理圖像，相對(duì)于CPU，可以同時(shí)處理大量相對(duì)簡(jiǎn)單的運(yùn)算，受控于CPU

TPU:?與TF軟件相關(guān)的神經(jīng)網(wǎng)絡(luò)學(xué)習(xí)的算法加速器

NPU：算法加速器的統(tǒng)稱（個(gè)人目前認(rèn)為TPU也是一種NPU）

DPU：主要用來處理“datacenter Tax”，承載網(wǎng)絡(luò)傳輸數(shù)據(jù)這一部分（不涉及算法訓(xùn)練）的算力

CPU

capabilities

Central processing unit,?a general-purpose processor, performs basic arithmetic, logic, controlling, and input/output (I/O) operations specified by the instructions in the program.

Specialized scenarios

Serial commputing

typical image

popular manufacturers

Intel, AMD, Qualcomm, NVIDIA, IBM, Samsung, Hewlett-Packard, VIA

GPU

capabilities

Graphic processing unit,?using for manipulating computer graphics and image processing, their highly parallel structure makes them more efficient than CPUs for algorithms that process large blocks of data in parallel. it's popular for machine learning, deep learning, GPUs allow certain calculations to be performed 10x to 100x faster than the same calculation running on a traditional CPU.

Specialized scenarios

Parallel? commputing

typical image

popular manufacturers

NVIDIA, AMD, Broadcom Limited

TPU

capabilities

Tensor processing unit,?an application-specific integrated circuit, to accelerate the AI calculations and algorithm. Google develops it specifically for neural network machine learning for the TensorFlow software. Google owns TensorFlow software.

Specialized scenarios

neural network machine learning for the TensorFlow software

typical image

popular manufacturers

google

NPU

capabilities

Neural network processing unit,?a class of specialized hardware accelerator or computer system designed to accelerate artificial intelligence and machine learning applications, including artificial neural networks and machine vision. Typical applications include algorithms for robotics, internet of things, and other data-intensive or sensor-driven tasks.

sample image

(Because it's a class of things, here is just a sample. The specific architecture depends on the specific task )

popular manufacturers

IBM, Intel, AMD,?google, NVIDIA

DPU

capabilities

Data processing unit(dataflow processing unit), specialists in moving data in data centers, DPUs, or data processing units, are a new class of programmable processor and will join CPUs and GPUs as one of the three pillars of computing.?

Specialized scenarios

Dealling with?"Datacenter Tax"(a set of shared low-level software components that comprise almost 30% of all processor cycles in production datacenters. The constituents of this "tax" -- the necessary components to do distributed computation (data serialization, compression, etc.)

typical image

popular manufacturers

NVIDIA, Intel, Broadoom, MS, Huawei

=END=

=references=

[1]https://serverguy.com/comparison/cpu-vs-gpu-vs-tpu/

[2]https://network.51cto.com/article/638974.html

[3]https://baike.baidu.com/item/DPU/11007035?fr=aladdin#2

[4]https://www.cambricon.com

[5]https://en.wikipedia.org/wiki/AI_accelerator

[6]https://pdf.dfcfw.com/pdf/H3_AP202107121503308793_1.pdf?1626099552000.pdf

[7]https://developer.nvidia.com/blog/embedded-machine-learning-cudnn-deep-neural-network-library-jetson-tk1/