Blackwell Breaks the 1,000 TPS/User Barrier With Meta’s Llama 4 Maverick

· NVIDIA Corporation · May 23, 2025, 12:36 a.m.
DGX Data Center / Cloud Generative AI LLMs artificial-intelligence NVIDIA large language models Performance Optimization
Summary
NVIDIA announces a groundbreaking achievement in large language model inference, showcasing that a single DGX B200 node with eight Blackwell GPUs can exceed 1,000 transactions per second (TPS) per user, setting a new record in performance for LLMs.
Read full post on developer.nvidia.com →
AUTHOR