DeepSeek V4 Pro 1.6T · Chip 对比

DeepSeek V4 Pro 1.6T — GB200 NVL72 vs MI355X

GB200 NVL72(NVIDIA Blackwell)与 MI355X(AMD CDNA 4)在 DeepSeek V4 Pro 1.6T 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。

DeepSeek V4 Pro 1.6T 在 46 tok/s/user 交互性下的吞吐量:GB200 NVL72 为 9562 tok/s/Chip,MI355X 为 3027。每百万 token 成本分别为 $0.05 和 $0.14。GB200 NVL72 每 token 成本低 156%;GB200 NVL72 每 Chip 吞吐量高出 216%。

GB200 NVL72 / MI355X 在 DeepSeek V4 Pro 1.6T 上以 75 tok/s/user 运行:8616 / 1613 tok/s/Chip,$0.06 / $0.26 每百万 token。GB200 NVL72 每 token 成本低 328%;GB200 NVL72 每 Chip 吞吐量高出 434%。

在 18–133 tok/s/user 交互性区间的高端,即 DeepSeek V4 Pro 1.6T 上以 104 tok/s/user 运行时:GB200 NVL72 达到 6714 tok/s/Chip($0.08/百万 token),MI355X 达到 777($0.53/百万)。GB200 NVL72 每 token 成本低 588%;GB200 NVL72 每 Chip 吞吐量高出 764%。 (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/chip)
GB200 NVL72:9562.2MI355X:3027.2
GB200 NVL72:8615.9MI355X:1613.3
GB200 NVL72:6714.0MI355X:777.4
Cost ($/M tok)
GB200 NVL72:$0.054MI355X:$0.138
GB200 NVL72:$0.060MI355X:$0.257
GB200 NVL72:$0.077MI355X:$0.531
tok/s/MW
GB200 NVL72:5113476MI355X:1448401
GB200 NVL72:4607427MI355X:771914
GB200 NVL72:3590381MI355X:371948
Concurrency
GB200 NVL72:~15421MI355X:~100
GB200 NVL72:~6302MI355X:~27
GB200 NVL72:~2196MI355X:~7

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:
部署模式:
投机解码: