gpu.md

GPU

Git hub: Nvidia/apex

• Automatic Mixed Precision (AMP)
  • to speed up net training, utilizing the Tensor Core microarchitecture
  • Tensor Core is enabled Volta, Turing microarchitecture, so not supported for Tesla Kepler series
• AMP is easy to use with 3 lines of code

NVIDIA GPU Cloud

• A hub for GPU optimized software
• Docker containers ready to be download and use

NVIDIA GPUs

• Desktop/Mobile/Worstation series
• Tesla is designed for worstation/server usage, without graphic output, only for parallel massive computing

GPU vs CPU

• CPU
  • optimised for sequencial serial processing
  • wide range of instructions
• GPU
  • initially for image rendering
  • Focus on small instruction set
  • Massive parallel computing, focus on throuput

Install GPU driver

• Official Installation Guide

• GPU Driver vs GPU CUDA Toolkit
  1）打开终端，先删除旧的驱动：

sudo apt-get purge nvidia*

2）Disable Nouveau

• nouveau is a open source gpu driver project for supporting nvidia drivers
• nouveau did not support CUDA.

sudo vim /etc/modprobe.d/blacklist-nouveau.conf
# add these line
blacklist nouveau
options nouveau modeset=0

# Regenerate the kernel initramfs
sudo update-initramfs -u

• Reboot

# reboot
sudo reboot

# check if Nouveau is killed
lsmod | grep nouveau

3）重启系统至init 3（文本模式），也可先进入图形桌面再运行init 3进入文本模式，再安装下载的驱动就无问题，

首先我们需要结束x-window的服务，否则驱动将无法正常安装

关闭X-Window，很简单：sudo service lightdm stop，然后切换tty1控制台：Ctrl+Alt+F1即可

4）接下来就是最关键的一步了：sudo./NVIDIA.run开始安装，安装过程比较快，根据提示选择即可最后安装完毕后，重新启动X-Window：sudo service lightdm start，然后Ctrl+Alt+F7进入图形界面；

• Links to Nvidia official gpu driver download page
• cudatoolkit may installed at

/usr/lib/x84_64-linux-gnu

# may find libcuda using
ldconfig -p |grep libcuda

NCCL

• installation guide