V100 SXM2 引脚定义与 NVLink 参考
V100 SXM2 pinout and NVLink reference

原始链接: https://github.com/pigeon-labs-stack/v100-sxm2-reference

本项目正在开发支持 NVLink 的 NVIDIA V100 SXM2 模块开源载板,面向两 GPU 和四 GPU 配置。目前它是一个公开参考套件,而非认证板:提供引脚映射、连接方案,以及针对一个、两个或四个模块的仅含连接器 KiCad 起始项目。参考文档记录了 800 个连接器触点的信号名、证据和电气说明,以及 NVLink 连接表。所有触点均已标注,但不包含稳压器、电源控制或时钟电路;需要外部时钟(156.25 MHz NVLink 和 100 MHz PCIe)。包含原理图、PDF、工作簿和预览,采用 Catppuccin Mocha 视觉风格。布线板、制造输出和认证仍为后续工作。欢迎在 CERN-OHL-W-2.0 和 CC BY-SA 4.0 下贡献与修正。

Hacker News 上的一个帖子介绍了一款 GitHub 上的 V100 SXM2 引脚定义与 NVLink 参考资料。评论者讨论了手动连接器映射、Facebook/Open Compute Big Basin/DGX-1 衍生原理图,以及 SXM2 转 PCIe 转接卡。多位用户指出 V100 仍可用于本地 LLM 推理,但已显老态:一名用户报告使用自定义引擎运行重度量化版的 DeepSeek V4.1,速度为 80–100 t/s,Qwen Flash Next 3.8 速度为 200–550 t/s;另一位建议使用 Q8/FP16 或推测解码。硬件细节包括 4 卡配置、25 Gbps RDMA 网络、张量/流水线并行以及 32GB 版本。其他人提醒 V100 最高支持 CUDA 12 和驱动 585,支持较为尴尬,而 GeForce 显卡因数据中心 EULA 限制可能更具性价比。总体而言,V100 价格实惠,但在现代 LLM 工作负载下受限。
相关文章

原文

V100 SXM2 reference kit

This project aims to become an open-source carrier-board design for NVIDIA V100 SXM2 modules, including NVLink connections for two- and four-GPU configurations.

While that board is being developed, this reference kit shares the pin mappings, connection plans and connector-only KiCad resources so other people can develop their own designs.

My contribution so far is primarily curation: gathering public information, comparing sources, reconciling connector orientation and signal names, and making the results easier to follow. The original reverse-engineering work belongs to its authors, credited in Sources and attribution.

  1. Download the complete reference ZIP from v0.1.0-reference, or use Code > Download ZIP for the current development files.
  2. Extract the archive and keep the directory structure. Choose a single-, two- or four-module KiCad starter, then open its .kicad_pro file in KiCad 10.0.7 or a compatible newer version. Keep the whole kicad/ folder so the shared symbol library is available.
  3. For reading without KiCad, use the complete schematic PDFs or standalone workbook.

The download guide includes print editions and library-path help. Change notes identify reference editions. This is a public reference pre-release, not a hardware-qualified carrier board.

  • 800 canonical contacts: 400 on connector A and 400 on connector B, with signal names, evidence and electrical-treatment notes.
  • Two- and four-GPU connection tables: exact Connector B pads, directions, lanes and polarity.
  • Three editable KiCad starters: one module’s A/B connectors, plus two- and four-module versions with NVLink data connected. Every contact stays labelled, including unconnected clocks, power and control pins. No regulator, power-control or clock circuits are included.
  • Catppuccin Mocha visuals: contact maps, topology diagrams, schematic previews and the standalone dark reference workbook. An optional KiCad colour theme is included.

External clock required: the NVLink starter schematics show data wiring, not a complete working NVLink implementation. Each V100 needs a 156.25 MHz differential NVHS clock on B/E20 (+) and B/F20 (−). The separate PCIe clock is 100 MHz on A/E7 (+) and A/F7 (−). Clock requirements and scope.

These are the actual connector-only KiCad schematics: 400 labelled contacts on A and 400 on B. Click a preview to open the full-size vector drawing. The two- and four-GPU NVLink arrangements are shown separately below and in the connection guides.

Connector A — PCIe, power contacts and control

Connector A KiCad schematic: all 400 labelled contacts

Connector B — NVLink and management

Connector B KiCad schematic: all 400 labelled contacts

Open the KiCad project · Schematic overview · KiCad guide

The full six-group two-GPU arrangement is derived from the documented pairing families; the four-GPU arrangement follows the curated topology. Each guide explains its source evidence and remaining uncertainties. These are logical connection plans, not routed bridge schematics.

Two-GPU logical NVLink group allocations

Open the two-V100 KiCad project · Native schematic previews

Four-GPU logical NVLink group allocations

Open the four-V100 KiCad project · Native schematic previews

Two-GPU connections · Four-GPU connections

Actual dual-connector footprint exported directly by KiCad

The actual local reference footprint, exported directly by KiCad: 800 contact pads, housing outlines and mounting features. A raw PCB footprint is not included in this edition; see footprint availability and open questions and visual provenance.

Connector geometry preview

CAD preview of the two connector arrays

This is a render of an existing project geometry preview, using the Amphenol connector model. It illustrates connector placement; it is not a completed carrier board or proof of connector A/B identity, mating-part selection or viewing direction. See render provenance.

The longer-term goal is a reusable open-source board design. This reference kit currently covers the pin mappings, NVLink connection plans and connector starter. Routed boards, manufacturing outputs and hardware qualification are future work. See the roadmap.

Corrections are particularly useful when they include an oriented drawing, exact source reference, continuity result or measured behaviour. Please distinguish a confirmed observation from an interpretation.

Use the correction and file-help forms, following the contribution guide. To reference this edition in your own work, use GitHub's Cite this repository option or CITATION.cff, and retain the relevant upstream source credits.

Our original hardware design contributions use CERN-OHL-W-2.0; our original documentation, diagrams and workbook presentation use CC BY-SA 4.0. Both allow commercial reuse, with attribution/notices and the applicable reciprocal sharing requirements. See licence scope and full texts and third-party notices. The licences do not claim ownership of pinout facts or replace upstream terms.

Follow the project and join the discussion on the Level1Techs forums: designing an open-source budget 4× V100 SXM2 NVLink backplane.

Thank you to everyone on the Level1Techs forums for your help, ideas and contributions to this project.

联系我们 contact @ memedata.com