NVIDIA Releases Blackwell B200 GPU Architecture with 20 PFLOPS FP4 Inference
NVIDIA announced the Blackwell B200 GPU featuring 208 billion transistors and second-generation Transformer Engine delivering 4x training and 30x inference speedups.
Verified State Diff
Comparison Mode:
- Previous State
Hopper H100 GPU delivered 4 PFLOPS FP8 inference with 80GB HBM3 memory.
+ Verified New State
Blackwell B200 delivers 20 PFLOPS FP4 inference with 192GB HBM3e memory at 8TB/s bandwidth.
Impact & Verification Analysis
WHO IS AFFECTED
AI researchers, cloud hyperscalers (AWS, Azure, GCP), and enterprise datacenter operators.
WHY IT MATTERS
Enables real-time inference on trillion-parameter frontier LLMs with 25x lower power consumption.
Full Fact Overview
Blackwell introduces 5th-generation NVLink interconnects with 1.8TB/s bidirectional throughput per GPU, designed specifically for trillion-parameter AI models.
Multi-Source Evidence Chain (1)
TRACKED ENTITY
Explore all historical NVIDIA changes