GPU, CUDA,cuDNN Summary of the relationship between the three

CPU and GPU

cpu and gpu Structure display
GPU（ image processor ,Graphics Processing Unit） and CPU（ a central processor ,Central Processing Unit） The main difference in design is GPU There are more units of operation （ Like the green one in the picture ALU）, and Control and Cache Unit is not as good as CPU many , This is because GPU In parallel computing, each computing unit executes the same program , It doesn't take too much control .Cache Cells are used for data caching ,CPU Can pass Cache To reduce the number of access to main memory , That is to reduce memory latency （memory latency）.GPU in Cache Very small or not , because GPU Memory latency can be reduced by parallel computing . therefore CPU Of Cahce The design is mainly to achieve low latency ,Control Mainly versatility , Complex logic control unit can guarantee CPU Distribute tasks and instructions efficiently . therefore CPU Good at logic control , It's serial computing , and GPU Good at high-intensity computing , It's parallel computing . For example ,GPU Like thousands of coolies , Everyone does the same kind of drudgery , There is no dependence on each other , They're all independent , Simple people are more powerful ;CPU It's like a contractor , Although I can also do coolie work , But there are fewer people , So they are generally responsible for task allocation , Personnel scheduling, etc .
It can be seen that GPU Acceleration is achieved by a large number of threads in parallel , So for work that cannot be highly parallelized ,GPU It doesn't work . and CPU It's a serial operation , It needs to be very versatile , It mainly plays the role of unified management and assignment of tasks .

CUDA

CUDA yes NVIDIA Launched for home GPU The framework of parallel computing , in other words CUDA Only in NVIDIA Of GPU Up operation , And only when the computing problem to be solved is a large number of parallel computing can play CUDA The role of .

cuDNN

cuDNN（CUDA Deep Neural Network library）： yes NVIDIA The acceleration library for deep neural network is built , It's a deep neural network GPU Acceleration Library . If you want to use GPU Training models ,cuDNN It's not necessary , But we usually use this acceleration library .

summary

Simply speaking ,CPU Suitable for serial computing , Good at logic control .GPU Good at high-intensity parallel computing , Apply to AI Algorithm training and learning
CUDA yes NVIDIA A framework dedicated to managing the allocation of computing units
cuDNN It is used for deep neural networks gpu Acceleration Library

Query version command ：

ubuntu:

cuda  edition  
nvcc -V
cat /usr/local/cuda/version.txt

cudnn  edition  
cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2

 see Nvidia GPU edition 
nvidia-smi

//10s Show once 
watch -n 10 nvidia-smi

nvidia-smi -l