GH200 vs H200

The FP16 dense figure of 989 TFLOPS against the FP32 figure of 67 TFLOPS yields the same ratio on both GPUs and therefore supports identical acceleration for mixed precision training and inference workloads. Memory bandwidth of 4800 GB/s on the H200 versus 4000 GB/s on the GH200 directly influences achievable batch sizes during large model execution. Dense FP16 performance remains 989 TFLOPS on each device and sparse FP16 performance remains 1979 TFLOPS on each device so throughput equivalence holds when sparsity is applied consistently.

Choose the GH200 when you rent a single GPU and the workload spills past 96 GB, because the NVLink-C2C link to the Grace CPU makes CPU side memory fast enough to hold overflow weights, KV cache, or large datasets. GPU compute is identical to the H200 at 989 TFLOPS dense FP16 and 1979 TFLOPS dense FP8, so you lose nothing on the accelerator itself. If you need several GPUs sharing NVLink in one node, the H200 in an SXM or NVL system is the standard option.

The H200 fits deployments that require 141 GB HBM3e memory and 4800 GB/s bandwidth at a 700W TDP. Its support for InfiniBand and NVL form factor suits scaled clusters that prioritize additional memory capacity and reduced power consumption over the GH200 configuration.