Why in news?
The US under President Donald Trump announced on December 8, 2025, the removal of export restrictions on Nvidia's H200 chips to China, its second-most powerful AI processor, while imposing a 25% fee on such sales. This marks a policy reversal amid US-China trade tensions, excluding Nvidia's top Blackwell and upcoming Rubin chips.
About Nvidia’s H200
Nvidia’s H200 is a high‑end data‑center GPU designed for generative AI and high‑performance computing, built on the Hopper architecture. It is essentially a memory‑supercharged successor to the H100 for large language models and other memory‑intensive workloads.​
Key specs and architecture
- The H200 is based on the Nvidia Hopper architecture and integrates 16,896 CUDA cores plus fourth‑generation Tensor Cores optimized for mixed‑precision AI (including FP8).
- It is the first GPU to use HBM3e, providing 141 GB of on‑package high‑bandwidth memory with up to 4.8 TB/s bandwidth, almost double the capacity and around 1.4× the bandwidth of H100.​
Performance and use cases
- For large language models and other transformer workloads, H200’s Transformer Engine and FP8 support can deliver several‑fold faster training versus older A100‑class GPUs and significantly faster inference versus H100, especially on very large models that are memory‑bound.
- The large HBM3e pool makes it well suited for generative AI, scientific simulations, and other HPC codes that need very high memory bandwidth and capacity, reducing the need for model or tensor sharding across many GPUs.​
Quick comparison with H100
| Feature |
Nvidia H100 |
Nvidia H200 |
| Architecture |
Hopper |
Hopper |
| HBM type |
HBM3 |
HBM3e |
| Memory capacity |
Up to 80 GB |
141 GB |
| Memory bandwidth |
~3.35 TB/s |
4.8 TB/s |
| Target workloads |
AI/HPC |
Larger, more memory‑intensive AI/HPC with LLM focus |
Download Pdf
