Wiki Schema

Domain

AI 基础设施:scale-up 网络、AI 加速器架构、确定性执行、推理系统、光学互联、LLM 模型架构、MoE、训练/推理系统、硬件-软件协同设计。

Conventions

  • 文件名:小写,连字符,无空格(如 nvidia-groq-3-lpx.md
  • 每个页面以 YAML frontmatter 开头
  • 使用 [[wikilinks]] 互相链接(每页至少 2 个出站链接)
  • 更新页面时更新 updated 日期
  • 新页面必须加到 index.md 对应分区
  • 所有操作追加到 log.md

Frontmatter

---
title: Page Title
created: YYYY-MM-DD
updated: YYYY-MM-DD
type: entity | concept | comparison | query | summary
tags: [from taxonomy below]
sources: [raw/articles/source-name.md]
---

Tag Taxonomy

  • 架构: accelerator, gpu, lpu, tpu, wse, chiplet, noc, interconnect, isa, pipeline, cache
  • 网络: scale-up, scale-out, ocs, tdm, wdm, routing, deterministic, fabric, transport, data-link, physical-layer, link-training, serdes, fec, retransmission, flow-control, congestion-control, switch, protocol, communication, now, network-on-wafer, ualink, ualoe
  • 推理: inference, decode, prefill, latency, throughput, serving, batching, disaggregated-inference, serving-system, kv-cache, reasoning, capacity-trap, parallelism
  • 编译器: compiler, scheduling, spatial-execution, deterministic-execution, programming-model
  • 公司: nvidia, cerebras, groq, amd, google, lightmatter, celestial-ai, deepseek, huawei, bytedance, graphcore, meta, samsung, sk-hynix, d-matrix, oxmiq, intel, microsoft, openai, broadcom, sambanova
  • 技术: sram, hbm, hbf, photonic, cpo, optical, mesh, memory-bandwidth, memory, rack, cpu, rpc, virtualization, power, 3d, tsv, monolithic, hybrid-bonding, through-silicon-via, microbump, cu-cu, packaging, integration, sequential-integration, wafer-on-wafer
  • AI系统: agentic-ai, ai-agent, moe, transformer, llm, training, attention, quantization, expert-parallelism, speculative-decoding, dataflow
  • 方法论: comparison, timeline, benchmark, architecture, information-theory, knowledge-management, formal-analysis
  • 模型: model, architecture, benchmark, training-system, inference-system
  • 技术: compression, sparse, optimization, routing, parallelism, kernel
  • 系统: sandbox, storage, hardware, networking, topology, infrastructure, interconnect, cxl, rdma

Rule: every tag on a page must appear in this taxonomy. If a new tag is needed, add it here first, then use it.

Page Thresholds

  • 创建页面:实体/概念在 2+ 来源出现,或对单一来源是核心主题
  • 更新已有页面:新来源提及已覆盖的内容
  • 不创建页面:掠过性提及、细节、领域外内容
  • 拆分页面:超过 ~200 行
  • 归档页面:内容完全被取代

Entity Pages

每个重要实体一页:概述、关键事实/日期、与其他实体的关系(wikilinks)、来源引用

Concept Pages

每个概念一页:定义/解释、当前认知状态、开放问题、相关概念

Update Policy

新信息与已有内容冲突时:

  1. 检查日期——较新来源通常优先
  2. 如果确实矛盾,记录两个立场及日期和来源
  3. 在 frontmatter 标记:contradictions: [page-name]