← Back to Intelligence

NVIDIA H200 Announced at SC23

Date: November 13, 2023
Company: NVIDIA
Category: Hardware & Infrastructure

Narrative

First GPU with HBM3e memory. 141GB capacity at 4.8 TB/s bandwidth. 76% more memory than H100. Nearly doubles Llama 2 70B inference speed.

NVIDIA

Reality

Shipped Q2 2024 as promised. Memory capacity gains verified. Inference speed improvements real but workload-dependent. Full H100 compatibility confirmed. Cloud providers deployed rapidly.

Implication

Mid-generation refresh validated memory-first optimization for AI inference. HBM3e became new standard. H200 bridged Hopper to Blackwell transition. Memory bandwidth competition intensified.

Tags

  • nvidia
  • gpu
  • inference