NVIDIA GPU 架构与产品线对应表

架构代号 (科学家) 发布年份 消费级（GeForce GTX(gaming)/RTX(AI with tensor core)） 专业图形（Quadro / RTX A） 数据中心（Tesla / A/H/B） 边缘AI（Jetson）
Fermi 2010 GTX 480 Quadro 6000 Tesla M2090 —
Kepler 2012 GTX 680 Quadro K6000 Tesla K80 Tegra K1
Maxwell 2014 GTX 980 Quadro M6000 — Jetson Nano
Pascal 2016 GTX 1080 Quadro P6000 Tesla P100 Jetson TX2
Volta 2017 — — Tesla V100 Jetson Xavier
Turing 2018 RTX 2080 Ti Quadro RTX 8000 T4 —
Ampere 2020 RTX 3080 / 3090 RTX A6000 A100 Jetson Orin 系列
Ada Lovelace 2022 RTX 4090 RTX 6000 Ada — —
Hopper Grace Hopper，编译器先驱（女性计算机科学家） 2022 — — H100 —
Blackwell David Blackwell，概率论与统计学家（非裔数学家） 2024 — — B100 / GB200 —

S - HW:SoC GPU

  • Discrete GPU（独立显卡） 独立于 CPU 的 GPU，通常是单独的芯片
  • SoC GPU（集成显卡）集成在 SoC中的 GPU
    • System on Chip: one chip has the system of CPU, GPU, and others.

T - Jetson devices (Jetson Orion Nano Developer Kit)

Device Jetson modules Jetson developer kits
HW Include CPU, GPU, memory, and limited storage, but ship without a carrier board or pre-installed software
- attach it to a carrier board designed or procured for your end product
- flash it with the software image you have developed		 includes a non-production-specification Jetson module attached to a reference carrier board
SW ships with no software preinstalled with NVIDIA JetPack SDK
Usecase suitable for deployment in a production environment not intended for production use, to develop and test only
Product Jetson Orin Nano module Jetson Orion Nano Developer Kit

A

https://developer.nvidia.com/buy-jetson?product=all&location=CN

S - SW:Linux + CUDA

T - Jetson software (L4T, jetpack SDK) alt text

Jetson Linux/L4T/Linux for Tegra

Jetpack SDK

  • Version
    • JetPack 6.2 via apt-show cache nvidia-jetpack
    • nvcr.io/nvidia/l4t-jetpack:r36.4.0
    • nvcr.io/nvidia/l4t-cuda:12.6.11-runtime cuda only
  • Include
    • Jetson Linux:36.4.3 via cat /etc/os-release
    • Jetson AI Stack: CUDA 12.6.10 via nvcc -V
    • Jetson Platform Services

Jetson Stats

A

1.Install Jetson Linux

# 1. Download the latest Jetson Linux release package and sample file system for your Jetson developer kit
# Jetson Linux release package
${L4T_RELEASE_PACKAGE} = Jetson_Linux_<version>_aarch64.tbz2

# Sample file system package
${SAMPLE_FS_PACKAGE} = Tegra_Linux_Sample-Root-Filesystem_<version>_aarch64.tbz2


# 2. Untar the files and assemble the rootfs
tar xf ${L4T_RELEASE_PACKAGE}
tar xpf ${SAMPLE_FS_PACKAGE} -C Linux_for_Tegra/rootfs/
cd Linux_for_Tegra/
./tools/l4t_flash_prerequisites.sh
./apply_binaries.sh


# 3. Put your Jetson developer kit into Force Recovery Mode
# Disconnect the power cable to ensure that the developer kit is powered off.
# Place a jumper to short the REC and GND pins on the 12-pin button header.
# Reconnect the power cable.
# Nvidia APX

# 4. Install/flash the Jetson release onto the Jetson developer kit
# jetson-orin-nano-devkit.conf is the .config file
./tools/kernel_flash/l4t_initrd_flash.sh --external-device nvme0n1p1 \
  -c tools/kernel_flash/flash_l4t_t234_nvme.xml -p "-c bootloader/generic/cfg/flash_t234_qspi.xml" \
  --showlogs --network usb0 jetson-orin-nano-devkit internal

# 5. Kit will auto reboot to Linux, remove the jumper

2.Config Jetson Linux

# 1. bluetooth
systemctl start bluetooth
systemctl enable bluetooth
bluetoothctl 
> discoverable on
> scan on
> pair MAC

# 2. Install broswer
snap install chromium

3.Install JetPack

apt update
apt install nvidia-jetpack

vim /etc/bash.bashrc
export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
export CUDA_ROOT=/usr/local/cuda

4.Install Jetson Stats

apt-get install python3-pip
pip3 install -U pip
pip3 install -U jetson-stats
systemctl restart jtop.service
systemctl enable jtop.service
jtop

S - PyTorch

T - PyTorch for Jetson Platform

  • Version
    • 2.0
    • nvcr.io/nvidia/l4t-pytorch:r35.2.1-pth2.0-py3

A

# 1. Install system packages required by PyTorch
apt-get -y update
apt-get install -y  python3-pip libopenblas-dev

# 2. 24.06 PyTorch or later versions, cusparselt needs to be installed first
wget raw.githubusercontent.com/pytorch/pytorch/5c6af2b583709f6176898c017424dc9981023c28/.ci/docker/common/install_cusparselt.sh 
export CUDA_VERSION=12.1 # as an example   
bash ./install_cusparselt.sh

# 3.install PyTorch 
export TORCH_INSTALL=https://developer.download.nvidia.cn/compute/redist/jp/v61/pytorch/torch-2.5.0a0+872d972e41.nv24.08.17622132-cp310-cp310-linux_aarch64.whl
# Specific version: https://developer.download.nvidia.com/compute/redist/jp/v$JP_VERSION/pytorch/$PYT_VERSION
# JP_VERSION: The major and minor version of JetPack you are using, such as 461 for JetPack 4.6.1 or 50 for JetPack 5.0.


python3 -m pip install numpy=='1.26.1'
python3 -m pip install --no-cache $TORCH_INSTALL
# Attempting uninstall: sympy
#     Found existing installation: sympy 1.9
#    ━━━━━━━━━━╺━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 2/8 [sympy]error: uninstall-distutils-installed-package

# × Cannot uninstall sympy 1.9
# ╰─> It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall.
--ignore-installed

# 4. Test Torch
import torch
# ImportError: libcusparseLt.so.0: cannot open shared object file: No such file or directory
# Install libcusparseLt: https://discuss.pytorch.org/t/importerror-libcusparselt-so-0-cannot-open-shared-object-file-no-such-file-or-directory/190061/2

pytorch.cuda.is_available()