Hardware for AI

Things on this page are fragmentary and immature notes/thoughts of the author. Please read with your own judgement!

TVM for deep learning is kind of like LLVM for programming languages.

Embedded/Edge AI¶

Jetson Nano

Google Coral

Intel Neural Compute Stick 2

https://heartbeat.fritz.ai/edge-tpu-google-coral-usb-accelerator-cf0d79c7ec56

https://blog.usejournal.com/google-coral-edge-tpu-vs-nvidia-jetson-nano-a-quick-deep-dive-into-edgeai-performance-bc7860b8d87a

https://mp.weixin.qq.com/s/NtnTQIecFq1L1ffPnirMIA

Tutorials¶

https://mp.weixin.qq.com/s/mId2Y4Do2k67pKq1WDVm8A

https://mp.weixin.qq.com/s/IraqkibhGciJ37sKoGxyvg

CPU：i9-10920X 显卡GPU：七彩虹RTX3090 Advance 内存：芝奇幻光戟16G x 4共64G 主板：华硕X299-DELUXE PRIME 固态硬盘：1TB西数NVME SSD + 1TB三星870QVO SATA SSD 机械硬盘：希捷EXOS 12TB氦气盘电源：海盗船AX1200i 1200W模组电源散热器：海盗船H100X240水冷 + 若干120机箱风扇机箱：海盗船AIR540 E-ATX机箱

最佳性价比GPU：RTX 2070
避免的坑：所有Tesla、Quadro、创始人版（Founders Edition）的显卡，还有Titan RTX、Titan V、Titan XP
高性价比：RTX 2070（高端），RTX 2060或GTX 1060 (6GB)（中低端）
穷人之选：GTX 1060 (6GB)
破产之选：GTX 1050 Ti（4GB），或者CPU（原型）+ AWS / TPU（训练），或者Colab
Kaggle竞赛：RTX 2070
计算机视觉或机器翻译研究人员：采用鼓风设计的GTX 2080 Ti，如果训练非常大的网络，请选择RTX Titans
NLP研究人员：RTX 2080 Ti
已经开始研究深度学习：RTX 2070起步，以后按需添置更多RTX 2070
尝试入门深度学习：GTX 1050 Ti（2GB或4GB显存）

https://www.digitalstorm.com/configurator.asp?id=3137888

Need UPS in case of power interruption

https://list.jd.com/Search?keyword=台式机断电保护电源&enc=utf-8&spm=2.1.11

References¶

Google Just Turned the Raspberry Pi into a Supercomputer