← Back to Intelligence

DGX GH200 AI Supercomputer

Date: May 29, 2023
Company: NVIDIA
Category: Hardware & Infrastructure

Narrative

256 Grace Hopper superchips. 1 exaflop AI performance. 144TB shared memory. NVLink creates single unified GPU.

NVIDIA

Reality

Deployed at Meta, Microsoft, Google. Exascale performance verified. NVLink fabric working but complex. Price estimated $50M+ per system. Long lead times.

Implication

Validated rack-scale thinking over individual GPU optimization. NVLink justified complexity premium. ARM Grace CPU proved viable in AI/HPC. Exascale training costs dropped significantly.

Tags

  • nvidia
  • compute
  • data-center
  • infrastructure