GLM 5/5.1 · GPU 对比

GLM 5/5.1 — B200 vs GB200 NVL72

B200NVIDIA Blackwell)与 GB200 NVL72NVIDIA Blackwell)在 GLM 5/5.1 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 GLM 5/5.1 上以 65 tok/s/user 运行时达到 2325 tok/s/GPU(每百万 token $0.23);GB200 NVL72 达到 6658 tok/s/GPU($0.09)。GB200 NVL72 每 token 成本低 152%;GB200 NVL72 每 GPU 吞吐量高出 186%。

GLM 5/5.1 在 91 tok/s/user 交互性下的吞吐量:B200 为 1692 tok/s/GPU,GB200 NVL72 为 2082。每百万 token 成本分别为 $0.32 和 $0.29。GB200 NVL72 每 token 成本低 9%;GB200 NVL72 每 GPU 吞吐量高出 23%。

B200 / GB200 NVL72 在 GLM 5/5.1 上以 116 tok/s/user 运行:1369 / 1199 tok/s/GPU,$0.40 / $0.51 每百万 token。B200 每 token 成本低 27%;B200 每 GPU 吞吐量高出 14%。 (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)
B200:2325.2GB200 NVL72:6658.4
B200:1691.9GB200 NVL72:2081.8
B200:1368.8GB200 NVL72:1199.3
Cost ($/M tok)
B200:$0.233GB200 NVL72:$0.093
B200:$0.321GB200 NVL72:$0.295
B200:$0.399GB200 NVL72:$0.506
tok/s/MW
B200:1359780GB200 NVL72:3560624
B200:989411GB200 NVL72:1113270
B200:800494GB200 NVL72:641314
Concurrency
B200:~17GB200 NVL72:~893
B200:~9GB200 NVL72:~82
B200:~5GB200 NVL72:~79

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: