NVIDIA H100 PCIe 80GB vs NVIDIA GH200 96GB HBM3

The GH200 is the newer part: Hopper, launched in 2023, against the H100's Hopper from 2022. Newer architectures typically add lower-precision formats and better throughput per watt, so check the precision rows your workload actually uses.

Dense throughput for the H100 against the GH200: FP64 26 vs 34 TFLOPS, FP32 51 vs 67 TFLOPS, FP16 757 vs 990 TFLOPS, FP8 1,513 vs 1,979 TFLOPS. Sparse figures, where the vendor publishes them, appear under each dense number in the performance table.

Memory is 80 GB against 96 GB, fed at 2.0 TB/s versus 4.0 TB/s. More memory per GPU means larger models fit before you have to shard across cards, which often matters more than raw TFLOPS for inference.

Power budgets are 350 W for the H100 and 1.0 kW for the GH200. At FP32 that works out to 0.15 against 0.07 TFLOPS per watt.

For LLM training and inference, weight the FP8 and FP16 rows and memory capacity most heavily. For scientific and HPC workloads, the FP64 row is the one to read.