MiniMax M3 428B — B300 vs H200
Head-to-head AI inference benchmark comparison of B300 (NVIDIA Blackwell) and H200 (NVIDIA Hopper) on MiniMax M3 428B. Latency, throughput, and cost across LLM workloads. Use the chart controls below to switch sequences, precisions, and metrics — same interactions as the main inference chart.
Setting 82 tok/s/user as the target on MiniMax M3 428B, B300 produces 2017 tok/s/chip ($0.31 per million tokens) and H200 produces 666 ($0.51). B300 is 61% cheaper per token; B300 delivers 203% more tok/s/chip.
At 153 tok/s/user interactivity on MiniMax M3 428B, B300 delivers 1031 tok/s/chip at $0.61 per million tokens; H200 delivers 331 tok/s/chip at $1.01. B300 is 66% cheaper per token; B300 delivers 212% more tok/s/chip at this point.
B300 posts 495 tok/s/chip for $1.26 per million tokens at 223 tok/s/user on MiniMax M3 428B; H200 posts 166 tok/s/chip for $2.10. B300 is 66% cheaper per token; B300 delivers 198% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B300:2016.9H200:666.5 | B300:1031.0H200:330.7 | B300:495.1H200:166.1 |
| Cost ($/M tok) | B300:$0.314H200:$0.505 | B300:$0.611H200:$1.011 | B300:$1.264H200:$2.102 |
| tok/s/MW | B300:1061549H200:486476 | B300:542654H200:241422 | B300:260569H200:121223 |
| Concurrency | B300:~68H200:~18 | B300:~14H200:~4 | B300:~8H200:~3 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.