对比
芯片、系统与模型的并排规格对比。每项数据均保留其来源标注——厂商声明与估算值并列时，均会明确标注。
| NVIDIA GB200 NVL72 The rack that is the computer: 72 GPUs, one NVLink domain, 120 kW, liquid only. | NVIDIA HGX H100/H200 The 8-GPU baseboard that defined the standard AI server. | |
|---|---|---|
| Overview — Configuration | 36 × GB200 superchips: 72 GPUs + 36 Grace CPUsVendor spec | 8 × H100/H200 SXM + 4 × NVSwitchVendor spec |
| Memory — HBM3E | 13.4 TB per rack (+ ~17 TB LPDDR5X)Vendor spec | — |
| Compute — FP4 inference | 1,440 PFLOPS (sparse, rack)Vendor spec | — |
| Overview — NVLink domain | All 72 GPUs at 1.8 TB/s each — 130 TB/s aggregateVendor spec | — |
| Power — Rack power | ~120 kW — liquid cooling requiredVendor spec | — |
| Overview — GPU-to-GPU | — | 900 GB/s NVLink 4, all-to-allVendor spec |
| Power — Typical server power | — | ~10 kW air-cooledVendor spec |
| Overview — Networking | — | Typically 8 × 400G ConnectX-7 + BlueFieldVendor spec |
"—"表示该记录未公布此项数据；不同厂商披露内容各异，我们不以猜测填补空白。