MiniMax M3 428B · GPU 对比

MiniMax M3 428B — B200 vs GB200 NVL72

B200NVIDIA Blackwell)与 GB200 NVL72NVIDIA Blackwell)在 MiniMax M3 428B 上的正面 AI 推理基准测试对比。涵盖各类 LLM 工作负载的延迟、吞吐量与成本。使用下方图表控件切换序列、精度和指标——交互方式与主推理图表相同。

B200 在 MiniMax M3 428B 上以 148 tok/s/user 运行时达到 4781 tok/s/GPU(每百万 token $0.11)。GB200 NVL72 在此工作点没有数据。

B200:2210 tok/s/GPU,每百万 token $0.25(MiniMax M3 428B 上以 290 tok/s/user 运行)。GB200 NVL72 在此点尚未测试。

在 MiniMax M3 428B 上以 432 tok/s/user 运行时,B200 吞吐量为 849 tok/s/GPU,每百万 token 成本 $0.64;GB200 NVL72 在此目标点没有基准测试数据。 (数据反映此 URL 的默认 8k/1k · fp4 选择——如果您在控件中更改序列、精度或模型,下方表格和图表会自动更新。)

查看每美元性能对比 →

Interpolated from real benchmark data. Edit target interactivity values below to compare at different operating points.
Metric
Interactivity (tok/s/user)
Interactivity (tok/s/user)
Interactivity (tok/s/user)
Throughput (tok/s/gpu)
B200:4780.7GB200 NVL72:
B200:2209.5GB200 NVL72:
B200:848.8GB200 NVL72:
Cost ($/M tok)
B200:$0.113GB200 NVL72:
B200:$0.245GB200 NVL72:
B200:$0.644GB200 NVL72:
tok/s/MW
B200:2795751GB200 NVL72:
B200:1292124GB200 NVL72:
B200:496395GB200 NVL72:
Concurrency
B200:~8GB200 NVL72:
B200:~4GB200 NVL72:
B200:~1GB200 NVL72:

推理性能

不同模型、硬件配置和服务参数下的推理性能指标。

厂商:
聚合模式:
投机解码:
B200 vs GB200 NVL72:MiniMax M3 推理基准测试 | InferenceX