---
格式版本: 2
标题: "Appendix B – Node Configurations — NVIDIA NVL72 AI Factory"
原文链接: "https://docs.nvidia.com/enterprise-reference-architectures/nvl72-ai-factory/latest/appendix-node-configurations.html"
发布日期: "2026-05-18"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "docbuild:last-update: May 18, 2026"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 6
发布时间严格候选数量: 3
发布时间原页读取状态: "原页面已读取"
发布时间未找到原因: ""
发布时间校准时间: "2026-07-24T18:37:25+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-06-15T16:12:19+08:00"
入库时间: "2026-07-24T10:37:25.305Z"
来源平台: "SerpApi Google"
搜索渠道: "serpapi"
搜索词: "Compute tray"
匹配关键词:
  - "Compute tray"
  - "NVL72"
相关厂家:
  - "NVIDIA"
相关专家:
  []
内容类型: "网页"
抓取工具: "XCrawl Scrape"
清洗工具: "XCrawl Markdown + LLM 正文裁剪（重试成功）"
原始附件:
  []
AI优质: "是"
AI打分: 95
AI分档: "高置信优质"
AI质检状态: "通过"
AI打分理由: "NVIDIA官方文档，详细列出GB300 NVL72的Compute tray配置、交换机、线缆规格，技术细节丰富，直接涉及超节点/AI Rack核心部件，来源权威，新颖性高。"
AI质检模型: "deepseek-v4-flash"
AI质检时间: "2026-07-24T19:48:36+08:00"
AI主题相关性: 20
AI来源权威性: 15
AI新颖性: 18
AI技术细节: 20
AI商业部署信号: 12
AI完整性: 10
采集批次: "2026年6月15日16点12分12秒"
采集批次ID: "20260615-161212-183"
去重键: "https://docs.nvidia.com/enterprise-reference-architectures/nvl72-ai-factory/latest/appendix-node-configurations.html"
---

# Appendix B – Node Configurations

This appendix provides information on the compute trays and networking configurations.

## Compute Trays

Table 10 shows the GB300 compute tray requirements.

**Table 10: GB300 compute node/tray requirements**

| Component | Configuration |
| --- | --- |
| CPUs | 2x NVIDIA Grace (C2), with total of 72 ARM Neoverse V2 cores connected with NVLink C2C |
| System Memory | 1 TB aggregated CPU LPDDR5 Main Memory |
| Boot Drive | 1 x 2 TB Gen4 M.2 for OS SSD |
| Data Cache Drive | 8 x 4 TB E1.S for Data Cache SSD |
| NVIDIA GPUs | 4x NVIDIA Blackwell Ultra GPU, with total of 720 GB of aggregated HBM3 memory |
| Compute Network Adapter | 2x NVIDIA Mezzanine Boards with 2x ConnectX-8 800 GB/s NICs in each one |
| Storage/in-band Network Adapter | 1x Dual-port QSFP112 NVIDIA BlueField-3 DPU operating at 400 GB/s |
| Platform Security | TPM 2.0, SHA-256 and support for secure boot |
| Baseboard Management Controller (BMC) | Host BMC with discrete connection, supports Redfish 1.4 or above, Serial-over-LAN via SSH, IPv6 support, support discovery of BMC IP in-band via Redfish |
| Host UEFI | UEFI 2.6 compliant BIOS implementation for HTTP(s) network boot |

## Switches and Cables

Table 11 and Table 12 provides an overview of the switches and cables used in the cluster. The network adapters of the GB300 NVL72, the storage servers, and the control plane servers are included for better understanding of the connectivity requirements.

**Table 11: Switching with required inter-switch links, transceivers, and cables**

| Component | 1 SU | 2 SUs | 3 SUs | 4 SUs | 5 SUs | 6 SUs | 7 SUs | 8 SUs |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| GB300 NVL72 Racks | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
| GB300 NVL72 Compute Trays | 18 | 36 | 54 | 72 | 90 | 108 | 126 | 144 |
| Grace Processors (two per tray) | 36 | 72 | 108 | 144 | 180 | 216 | 252 | 288 |
| Blackwell GPUs (four per tray) | 72 | 144 | 216 | 288 | 360 | 432 | 504 | 576 |
| ConnectX-8 SuperNIC, Compute fabric (four per tray) | 72 | 144 | 216 | 288 | 360 | 432 | 504 | 576 |
| BlueField-3 B3240 DPU, In-band management, Customer & Storage fabric and Support Servers (one per tray) | 18 | 36 | 54 | 72 | 90 | 108 | 126 | 144 |
| SN5600 Ethernet switch, compute core fabric | 12 | 12 | 24 | 24 | 32 | 32 | 44 | 44 |
| SN5600 Ethernet switch, converged core fabric | 2 | 2 | 5 | 6 | 9 | 10 | 11 | 11 |
| SN2201 leaf switches for OOB management fabric | 2 | 4 | 6 | 8 | 10 | 12 | 14 | 16 |
| Twin-port OSFP, 2x400 Gbps Multimode SR8, 50m. Used for switch-to-switch connections (N/S Network) | 20 | 30 | 192 | 256 | 320 | 384 | 448 | 512 |
| MPO/APC to MPO/APC Multi-mode fibers. Switch-to-switch connections (N/S network). Length encoded by xx. | 20 | 30 | 192 | 256 | 320 | 384 | 448 | 512 |
| Twin-port OSFP, 2x400 Gbps Multimode SR8, 50m. Used for switch-to-switch connections (E/W Network) | 288 | 288 | 576 | 576 | 864 | 864 | 1152 | 1152 |
| MPO/APC to MPO/APC Multi-mode fibers. Switch-to-switch connections (E/W network). Length encoded by xx. | 288 | 288 | 576 | 576 | 864 | 864 | 1152 | 1152 |
| 100G optical modules for OOB leaf switches, transceiver, 100GbE, QSFP28, MPO, 1301nm, DR1, up to 100m | 4 | 8 | 12 | 16 | 20 | 24 | 28 | 32 |
| Twin-port 2x400G optical transceivers for spine switches, 2xDR4, single-mode, OSFP, 2x MPO/APC optical connectors | 2 | 2 | 2 | 2 | 4 | 4 | 4 | 4 |
| MPO/APC to LC Single-mode fibers. OOB MgmtSW-to-switch. One to Four breakout. Length encoded by xx. | 2 | 2 | 4 | 4 | 6 | 6 | 8 | 8 |

**Table 12: Endpoint connects with required transceivers, and cables**

| Component | 1 SU | 2 SUs | 3 SUs | 4 SUs | 5 SUs | 6 SUs | 7 SUs | 8 SUs |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| GB300 NVL72 Racks | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 |
| GB300 NVL72 Compute Trays | 18 | 36 | 54 | 72 | 90 | 108 | 126 | 144 |
| Grace Processors (two per tray) | 36 | 72 | 108 | 144 | 180 | 216 | 252 | 288 |
| Blackwell GPUs (four per tray) | 72 | 144 | 216 | 288 | 360 | 432 | 504 | 576 |
| ConnectX-8 SuperNIC, Compute fabric (four per tray) | 72 | 144 | 216 | 288 | 360 | 432 | 504 | 576 |
| BlueField-3 B3240 DPU, In-band management, Customer & Storage fabric and Support Servers (one per tray) | 18 | 36 | 54 | 72 | 90 | 108 | 126 | 144 |
| Twin-port OSFP, 2x 400 Gbps Multimode SR8, 50m. Used in switch ports for Compute Node SuperNIC-to-switch connections | 72 | 144 | 216 | 288 | 360 | 432 | 504 | 576 |
| Twin-port OSFP, 2x 400 Gbps Multimode SR8, 50m. Used in switch ports for Compute Node DPU-to-switch connections | 18 | 36 | 54 | 72 | 90 | 108 | 126 | 144 |
| Twin-port OSFP, 2x 400 Gbps Multimode SR8, 50m, used in Compute Node SuperNIC for server-to-switch connections | 72 | 144 | 216 | 288 | 360 | 432 | 504 | 576 |
| Single-port 400 Gbps optical transceivers, multi-mode, QSFP112, used in Compute Node DPU ports for server-to-switch connections | 36 | 72 | 108 | 144 | 180 | 216 | 252 | 288 |
| MPO/APC to MPO/APC Multi-mode fibers. ConnectX-8 SuperNIC-to-switch. Length encoded by xx. | 144 | 288 | 432 | 576 | 720 | 864 | 1008 | 1152 |
| MPO/APC to MPO/APC Multi-mode fibers. DPU B3240-to-switch. Length encoded by xx. | 36 | 72 | 108 | 144 | 180 | 216 | 252 | 288 |
| Twin-port 2x400G optical transceivers for spine switches, 2xDR4, single-mode. Used in switch ports for storage-to-switch connections | 2 | 4 | 4 | 6 | 6 | 8 | 8 | 10 |
| 100G optical transceiver, 100GbE, QSFP28, MPO, 1301nm, DR1, up to 100m. Used in upstream storage for 100Gb QSFP28 NRZ connectivity | 10 | 18 | 28 | 36 | 46 | 54 | 64 | 72 |
| MPO/APC to LC Single-mode fibers. Storage-to-switch. One to Four breakout. Length encoded by xx. | 4 | 6 | 8 | 10 | 12 | 14 | 16 | 18 |
| Twin-port 2x400G optical transceivers for spine switches, 2xDR4, single-mode. Used in switch ports for customer network-to-switch connections | 4 | 6 | 8 | 10 | 12 | 14 | 16 | 18 |
| 100G optical transceiver, 100GbE, QSFP28, MPO, 1301nm, DR1, up to 100m. Used in upstream customer network for 100Gb QSFP28 NRZ connectivity | 18 | 36 | 54 | 72 | 90 | 108 | 126 | 144 |
| MPO/APC to LC Single-mode fibers. Customer network-to-switch. One to Four breakout. Length encoded by xx. | 6 | 10 | 14 | 18 | 24 | 28 | 32 | 36 |
| Twin-port OSFP, 2x 400 Gbps Multimode SR8, 50m. Used in switch ports for support servers DPU-to-switch connections | 12 | 12 | 12 | 12 | 12 | 12 | 12 | 12 |
| Single-port 400 Gbps optical transceivers, multi-mode, QSFP112, used in support server DPU ports for server-to-switch connections | 48 | 48 | 48 | 48 | 48 | 48 | 48 | 48 |
| MPO/APC to 2xMPO/APC breakout multi-mode fibers. Support server ConnectX-8 400Gb (single port) to switch. Length encoded by xx. | 24 | 24 | 24 | 24 | 24 | 24 | 24 | 24 |
| CAT6 RJ45 cable for 1G OOB fabric connections | 71 | 142 | 213 | 284 | 355 | 426 | 497 | 568 |

For exact BOM and part numbers, please work with your OEM/ODM partners and our NVIDIA support team depending on your specific rack and data center conditions.