MiniMax M2.5/M2.7 — B300 vs H200
Head-to-head AI inference benchmark comparison of B300 (NVIDIA Blackwell) and H200 (NVIDIA Hopper) on MiniMax M2.5/M2.7. 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.
Throughput at 52 tok/s/user on MiniMax M2.5/M2.7: B300 hits 4734 tok/s/chip, H200 hits 1095. Per-million costs land at $0.13 and $0.32 respectively. B300 is 138% cheaper per token; B300 delivers 332% more tok/s/chip.
B300 / H200 on MiniMax M2.5/M2.7 at 80 tok/s/user: 1504 / 541 tok/s/chip, $0.42 / $0.58 per million tokens. B300 is 38% cheaper per token; B300 delivers 178% more tok/s/chip.
Toward the upper edge of the 25–134 tok/s/user interactivity band, at 107 tok/s/user on MiniMax M2.5/M2.7: B300 runs 903 tok/s/chip at $0.69/M tokens, H200 runs 303 at $1.11/M. B300 is 61% cheaper per token; B300 delivers 198% more tok/s/chip. (Numbers reflect the default 1k/1k · fp8 selection for this URL — table and chart below update if you change sequence, precision, or model in the controls.)
| Metric | Interactivity (tok/s/user) | Interactivity (tok/s/user) | Interactivity (tok/s/user) |
|---|---|---|---|
| Throughput (tok/s/chip) | B300:4734.2H200:1095.1 | B300:1503.7H200:540.6 | B300:902.7H200:303.3 |
| Cost ($/M tok) | B300:$0.133H200:$0.317 | B300:$0.422H200:$0.582 | B300:$0.687H200:$1.108 |
| tok/s/MW | B300:2491677H200:799362 | B300:791431H200:394579 | B300:475099H200:221394 |
| Concurrency | B300:~478H200:~42 | B300:~42H200:~15 | B300:~17H200:~6 |
Inference Performance
Inference performance metrics across different models, hardware configurations, and serving parameters.