[NOTE] Nvidia GPU, CUDA and PTX
AI hardware accelerators, including FPGA, and other emerging ASICs like Google TPU, AMD AIE, AWS Trainium, etc. This note does not include details about the software stack, but only focuses on the hardware architecture and its design choices.
Hopper H100 Architecture
- PH
Basics
- SXM (Server PCI Express Module) is a form factor (for high bandwidth socketed solution) for Nvidia GPUs to connect to the host CPU. It is different from standard PCIe cards, and it is used in DGX/HGX servers for higher bandwidth and lower latency. SXM cards can interconnect with NVSwitch.