GB300 vs H200

The FP16 dense performance of the GB300 reaches 2250 TFLOPS compared with 989 TFLOPS FP16 dense on the H200 which produces a ratio of 2250 to 989. This difference directly scales training throughput for large models and supports larger batch sizes during inference when memory bandwidth remains the limiting factor. The GB300 memory bandwidth of 8000 GB/s versus 4800 GB/s on the H200 permits correspondingly larger batch sizes before memory capacity or transfer rates become constraints. FP16 with sparsity follows the same ratio at 4500 TFLOPS versus 1979 TFLOPS while FP8 dense figures stand at 4500 TFLOPS versus 1979 TFLOPS. FP32 performance shows a narrower gap at 75 TFLOPS versus 67 TFLOPS which affects precision-sensitive scientific workloads less dramatically than the low-precision tensor operations.

The GB300 suits LLM training workloads that require the full 262-288 GB HBM3e capacity and 2250 TFLOPS FP16 dense throughput. Its 8000 GB/s memory bandwidth further enables larger batch sizes than the 4800 GB/s available on the H200. Organizations running FP8 dense operations at 4500 TFLOPS also select the GB300 when maximum tensor throughput is required.

The H200 fits deployments where the 700 W TDP must remain within existing power envelopes that cannot accommodate the 1400 W TDP of the GB300. Its 141 GB HBM3e capacity and 1979 TFLOPS FP16 dense performance remain sufficient for inference tasks that do not exceed those memory or throughput limits.