DGX H100 vs DGX GH200 256 GPU SuperPod Memory Common Address Space

Accelerated Computing DGX User Forum
1

My understanding is that the DGX H100 SuperPods only share a common address space among the 8GPUs within a BasePod. Is this the same with the DGX GH200 systems? The language in the DGX GH200 tech blog ( NVIDIA DGX GH200 Tech Blog) seems to indicate that the DGX GH200 has a full common shared address space for all 256 GH200 devices. If this is the case, can we assume that the NVLink “network” is not needed for DGX GH200, as NVLink Network endpoints do not share common memory address space.

Thanks in advance!

2

Excellent question @highlander.email.server , and glad to see you’re interested in the new DGX systems using our Grace chip!

I encourage you to come to GTC this year (GTC 2024: #1 AI Conference ) where we’ll be able to share more information about how these systems are designed, operated, and look at to applications.

ScottE