NVIDIA H20 96GB vs NVIDIA H200 SXM 141GB

Both GPUs launched in 2024: the H20 on Hopper and the H200 on Hopper.

Dense throughput for the H20 against the H200: FP64 1 vs 34 TFLOPS, FP32 44 vs 67 TFLOPS, FP16 148 vs 990 TFLOPS, FP8 296 vs 1,979 TFLOPS. Sparse figures, where the vendor publishes them, appear under each dense number in the performance table.

Memory is 96 GB against 141 GB, fed at 4.0 TB/s versus 4.8 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 400 W for the H20 and 700 W for the H200. At FP32 that works out to 0.11 against 0.10 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.