INFERENCE KING
Hopper Data Center H200 SXM Hopper
NVIDIA H200 SXM5
Same Hopper compute power as the H100 — with 141 GB of next-gen HBM3e and 4.8 TB/s of memory bandwidth. The H200 SXM5 unlocks larger model inference and dramatically higher throughput for trillion-parameter LLMs that simply cannot fit in smaller memory envelopes.
16,896 CUDA Cores
141 GB HBM3e
4800 GB/s Bandwidth
700W TDP
MSRP $35,000
Architecture Hopper
GPU Chip GH100
Fabrication TSMC 4N
CUDA Cores 16,896
SM Count 132
Tensor Cores 528
RT Cores 0
Boost Clock 1,980 MHz
Memory Size 141 GB HBM3e
Memory Bus 6144-bit
Memory Bandwidth 4800 GB/s
L2 Cache 51 MB
TDP 700 W
Launch Date March 18, 2024
- 141 GB HBM3e Memory
- 4.8 TB/s Memory Bandwidth
- 4th-Gen Tensor Cores (FP8)
- Transformer Engine
- NVLink 4.0 (900 GB/s)
- Multi-Instance GPU (MIG)
- Confidential Computing
Trillion-Parameter LLM Inference
AI Training
Mixture-of-Experts Models
Drug Discovery
HPC