62 · 01

Red Flags

These thoughts mean STOP—you're rationalizing:

| Thought | Reality |
|---------|---------|
| "This is just a simple question" | Questions are tasks. Check for skills. |
| "I need more context first" | Skill check comes BEFORE clarifying questions. |
| "Let me explore the codeba
succeeded in 0ms:
---
name: product-exploded-view-company-prompts
description: Use when the user requests prompt-only product exploded-view, teardown, component callout, hardware decomposition, industry-chain company-name recommendations, Chinese 拆解图/爆炸图 poster prompts, or prompt-ready Chinese poster copy without generating or editing images.
---

Product Exploded View Company Prompts

Overview

Create ready-to-copy image-generation prompts for product exploded-view posters without generating images. Preserve the source skill's teardown structure and optionally add relevant company names for an industry chain. Add no company information beyond the names.

Bundled Resources

- `references/archive/`: historical snapshots only; do not load unless the user asks to inspect an older version.
- `scripts/validate_output.py`: deterministic validator for an existing prompt/copy draft; run it from a file or standard input when mechanical compliance checks are useful.
- `scripts/lint_skill.py`: static linter for this skill folder; run it after editing the skill.
- `references/geographic-terms.txt`: deterministic denylist for geographic and regional text; do not quote its entries in output.
- `tests/`: validator regression tests; use them only when maintaining this skill.

Boundary

- Output text only. Do not call `imagegen`, Imagen, browser rendering, SVG/HTML/canvas, Pillow, slides, screenshots, or any other image creation/editing tool.
- Do not create output bundle folders, image files, verification files, transparent padding, or post-processing scripts.
- Bundled text validators are allowed. They only read draft text, print diagnostics, and must not create image assets or persistent verification artifacts.
- Do not claim that an image exists, has a ratio, contains visible text, or passed visual inspection.
- If the user later asks to generate, edit, pad, or verify an actual image, switch to the appropriate image/poster workflow instead of this prompt-only skill.
- For premium multi-view engineering infographics focused on specifications, performance, orthographic views, cutaways, or dashboards, use `premium-engineering-infographic-prompts` instead of adding that workflow here.
- For industrial packaging illustration sheets, dieline concepts, fold/seam studies, and packaging material callouts, use `industrial-packaging-design-prompts` instead of adding that workflow here.
- Add only company names. Do not add company introductions, products, rankings, reasons, links, stock codes, locations, or other company data.
- Never output geographic or regional names.
- Never output company logos, logo-like marks, icons, or brand identity elements.

Workflow

1. Identify the subject, target language, style, audience/platform, and any required account credit from the user request.
2. Fix every poster prompt to an exact `3:4` portrait aspect ratio. Do not allow `2:3`, `4:5`, `9:16`, or any other ratio.
3. For real products, verify component names from official or primary sources when accuracy matters. If not verified, phrase the prompt as a "conceptual structure diagram", "schematic exploded view", "示意结构图", or "概念爆炸图".
4. If the user does not provide an author/account credit, infer from thread context. For this user's Chinese content workflow, default to `作者：好用工具推荐`.
5. Add a source disclaimer to every Chinese technical poster prompt and to the end of every individual publish-ready platform version. Unless the user supplies different wording, use `声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。` Never replace the per-platform endings with one shared, reusable, or "通用结尾" disclaimer. Require it as visible lower-middle image text near the author credit, but do not claim it was rendered or inspected.
6. Choose a visual style explicitly. Treat dark background, light background, blueprint, studio render, data-center hardware, and editorial infographic as optional styles, not fixed defaults. User-specified style wins; otherwise choose a topic-appropriate style and write it into the final prompt.
7. Build the prompt around this reading path:
`complete product -> decomposition guide axis -> numbered component layers`.
Honor an explicit user layout request first. Otherwise use the center-hero dual-side preset by default. Use a non-dual layout only for an allowed clarity exception defined below.
8. Prefer 7-10 numbered layers. Group uncertain or tiny internals into meaningful systems such as compute, memory, power, cooling, frame, connector, enclosure, sensor, optics, battery, interface, package, interposer, or control layer.
9. Keep all visible text in the requested language. For Chinese posts, require simplified Chinese, modern sans-serif typography, short labels, and no Japanese/ghost text.
10. Add a lightweight text-risk note for the prompt and generated publishing copy. Flag politics, sexual/adult content, graphic violence, gambling, drugs, hate/extremism, illegal activity, and other obvious platform-risk terms. Treat technical false positives separately, for example `爆炸图/爆炸展开` as exploded-view terms, `安全隔离` as a security/networking term, or `确定性` as a performance term.
11. For branded products, require no obvious brand logos, official logo glyphs, trademark emblems, or logo-like marks in the image. Use neutral product markings and text-only labels when product naming is needed.
12. Make poster titles attractive but editorial, not promotional. Prefer curiosity, contrast, or structural insight over launch-ad language, superlatives, or sales claims.
13. For Chinese poster requests, generate publishing copy by default unless the user explicitly asks for prompt-only output or no social copy.
14. Optionally make in-image content feel richer by adding secondary labels, compact annotation panels, tiny spec tags, and small explanatory notes, while preserving hierarchy, readability, and non-overlap.
15. When the subject includes an industry chain, select relevant publicly traded companies for each component or process stage. Verify relevance and listing status from an exchange, issuer investor-relations page, or another primary source when accuracy matters; never expose the research or classification.
16. Attach company names to the corresponding numbered component annotation. Add names to every relevant module when space and factual relevance allow; do not collect them elsewhere.
17. Internal precedence: A-share-related companies, then Hong Kong-listed companies, then other publicly traded companies. Apply this precedence independently inside each module. Treat a dual-listed company with an A-share relationship as first-class; otherwise use the second or third class. Exclude unlisted companies.
18. Never expose or label the market precedence. Do not create headings, group names, blank-line groups, columns, or explanatory notes that reveal it.
19. Use identical typography, font size, color, spacing, and separators for every company name. Do not translate, transliterate, regroup, or restyle names to signal categories.
20. Exclude any candidate whose displayed company name contains a geographic or regional term. Do not abbreviate, rewrite, or truncate a legal name to bypass this rule.
21. Output company names only. Do not output a logo, icon, introduction, product, location, link, stock code, ranking, or recommendation reason.
22. When a complete draft is already available through a file or standard input, run `scripts/validate_output.py` in the matching mode and fix deterministic errors before returning it. Do not create a temporary verification file solely to run the validator.

Text Validation

Run the output validator only when the draft is already available as a file or standard input:

bash
python3 scripts/validate_output.py draft.md --mode prompt-only
python3 scripts/validate_output.py draft.md --mode social
python3 scripts/validate_output.py draft.md --mode full-comments
python3 scripts/validate_output.py - --mode social < draft.md

The validator reports deterministic errors and advisory warnings. It scans every output line for configured geographic or regional terms, rejects visible market/language grouping, and validates component-attached company clauses. It cannot prove listing status or market precedence from bare names; verify those facts from primary sources. When editing this skill itself, run `python3 scripts/lint_skill.py` and `python3 -m unittest discover -s tests -v` from the skill directory.

Output Format

Return:

1. A short assumptions line when any product internals, style, or author credit were inferred.
2. One final prompt in a fenced code block, ready to paste into an image-generation model.
3. Optional prompt variants only when the user asks for alternatives.
4. For Chinese posters, include a `发布文案` section by default unless the user asks for prompt-only output or no social copy. The section must include 30 title candidates, mark one as recommended, include a separate `互动提问` section, and append the full source disclaimer directly to every individual platform version.
5. Every `互动提问` section must contain exactly 10 questions. Do not increase or decrease the count for a comment pack or any other mode.
6. When company recommendations are relevant, append bare names directly to the corresponding numbered component annotations.
7. A short "not generated" note when helpful: state that no image was created or inspected.

Company Name Recommendations

Only output bare company names.

- Select names relevant to the depicted product, component, process, or industry-chain role.
- Attach company names to the corresponding numbered component annotation.
- Internal precedence: A-share-related companies, then Hong Kong-listed companies, then other publicly traded companies.
- Apply the precedence independently within each component. A dual-listed company with an A-share relationship takes the first class; otherwise use the second or third class. Exclude unlisted companies.
- Never expose or label the market precedence.
- Use identical typography, font size, color, spacing, and separators for every company name.
- Do not translate, transliterate, regroup, or restyle names to signal categories.
- Do not create a standalone company panel, list, legend, footer, side column, or detached callout.
- After `相关公司：`, use `、` as the only separator and include no bullets, numbering, subheadings, brackets, classifications, or explanations.
- Omit any name containing a term from `references/geographic-terms.txt` or an identifiable country, territory, province, state, city, prefecture, county, district, or other place name.
- Do not shorten or alter a company name to remove a prohibited geographic term.
- Do not append a company introduction, product, business, location, ranking, reason, URL, stock code, or other metadata.
- Never output geographic or regional names.
- Never output company logos, logo-like marks, icons, or brand identity elements.

Use this exact machine-checkable form inside each relevant numbered component line. Replace the bracketed values with verified content:

text
01 "[component]" - [visual description] | 相关公司：[name]、[name]、[name]

Aspect Ratio Rules

- Every poster prompt must use an exact `3:4` portrait aspect ratio.
- Do not permit `2:3`, `4:5`, `9:16`, landscape, square, or any other ratio.
- Keep all content inside a safe area with approximately `5%` margin on every side so titles, labels, callouts, author credit, and source disclaimer do not touch or cross the canvas edge.
- Begin every final prompt and every prompt variant with the four mandatory lines below, verbatim and before any other prompt text:

text
画布比例必须为 3:4 竖版（Portrait 3:4）。
不允许生成 2:3、4:5、9:16 或其他比例。
所有内容需完整落在安全区域内，�
codex
我看到了 skill 的关键要求：这里只输出提示词与文案，不生成图片；它默认会加作者署名，但你的强约束明确覆盖这一点，所以最终内容会完全去掉署名和账号名。
exec
/bin/zsh -lc "sed -n '260,620p' /Users/alex/.codex/skills/product-exploded-view-company-prompts/SKILL.md 2>&1 | head -c 12000" in /Users/alex/.openclaw/workspace/exploded-poster-gen
succeeded in 0ms:
Left side:
"[heading]\n[short explanatory sentence]"
Right side:
"[heading]\n[short explanatory sentence]"

Bottom:
- Headline: "[bottom headline]"
- Body: "[2-3 concise sentences]"
- Footer mark: "[category / use case]"

Avoid:
garbled text, wrong language, Japanese text in Chinese posters, obvious brand logos, official logo glyphs, logo-like marks, company icons, company introductions, company products, company locations, company links, stock codes, rankings, recommendation reasons, geographic or regional names, origin-group labels, hard-ad title, sales CTA, exaggerated superlatives, hype words such as 重磅/颠覆/神器/必看/最强/炸裂/王炸, missing author credit, missing source disclaimer, metadata-only credit or disclaimer, watermark treatment, decorative clutter, empty-looking layout, unreadable tiny text, cropped labels, overlapping microtext, overlapping callouts, isolated floating parts without whole-to-parts guidance, unsupported exact claims about proprietary internals.

Component Rules

- Label every layer with a visible number and a short component name.
- Prefer 7-10 layers. Fewer can feel shallow; more often becomes unreadable.
- Group tiny internals into meaningful systems: compute, memory, power, cooling, frame, connector, enclosure, sensor, optics, battery, interface.
- Use callouts to explain function, not repeat component names.
- Keep primary image-model text short. Prefer one heading and one short sentence per main callout.
- Use "示意", "结构概念", "conceptual", or "schematic" for uncertain internals.
- For architecture or chips, describe whole package, stacked layers, key interconnect, base/control layer, package/interposer, and what the structure enables.
- For consumer products, describe shell/enclosure, display/interface, compute/control board, battery/power, sensors, connectors, thermal path, frame, and fastening when relevant.

In-Image Density Rules

- Use richer in-image text as an optional enhancement for technical posters when it helps the image feel complete and educational.
- Add 4-8 secondary details such as micro labels, small spec chips, short system-flow captions, mini legends, material notes, or compact annotation panels.
- Use smaller typography for secondary details, but do not make any required label, number, author credit, source disclaimer, or main callout tiny.
- Richer text should support the visual reading path, not become decorative filler.
- Avoid dense paragraphs inside the image; prefer short fragments, labels, and one-line notes.

Title Rules

- Poster titles should feel like technical editorial hooks, not product ads.
- Prefer titles that reveal a question, contrast, or structural role, such as "GPU 背后的调度底座", "为什么它不是显卡", or "AI 工厂里被低估的 CPU".
- Avoid hard-marketing title patterns such as "重磅发布", "颠覆未来", "最强芯片", "必看神器", "王炸新品", and sales-style calls to action.
- Keep the title shorter than the subtitle; let the subtitle carry the precise product name or use case when the title is more editorial.
- For Chinese publishing copy, generate 30 distinct title candidates by default and mark one recommended title.

Author Credit Rules

- Every prompt must include an in-image author/account credit.
- For Chinese content, use `作者：{name}`.
- Default to `作者：好用工具推荐` when no conflicting account appears.
- Place the credit inside the lower-middle content area, beside the lower decomposition axis, next to a real callout group, or inside a small engineering annotation panel connected to the diagram.
- Make the credit feel naturally merged with the real content so it is not easy to remove by cropping the bottom or deleting a detached footer.
- Avoid the very bottom edge, detached footer strips, corner-only labels, and large watermark treatment. Also avoid placing it too high or over the main reading path.
- Keep the credit smaller than component labels but clearly readable.
- Do not let it cover the product, numbered layers, arrows, or callouts.
- Do not treat the credit as a watermark outside the design; integrate it as a technical poster annotation.
- Do not use placeholders such as `【作者名】`, `{author}`, `[account]`, or `{name}` in the final prompt.

Source Disclaimer Rules

- Every Chinese technical poster prompt and companion publishing copy must include a source disclaimer.
- Default wording: `声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。`
- If the user supplies different wording, preserve it unless it creates a factual or legal misrepresentation.
- Require the disclaimer as visible image text near the author credit or in a connected lower-middle technical annotation panel. Keep it smaller than primary labels but clearly readable.
- Do not place it at the bottom edge, in a detachable footer, only in metadata, or outside the image.
- Append the full disclaimer directly at the end of every publish-ready platform version, including 作者, 小红书, 朋友圈/社群, and 知乎/B站动态.
- Do not consolidate, deduplicate, or replace these per-platform endings with a shared disclaimer, reusable ending, "通用结尾", footnote, or instruction to add it later.
- Do not claim that the generated image contains or correctly renders the disclaimer because this skill does not generate or inspect images.
- The disclaimer does not replace source verification. Continue to use official or primary sources when accuracy matters and label uncertain internals as conceptual or schematic.

Chinese Social Copy

Include social-media publishing copy by default for Chinese poster prompts unless the user asks for prompt-only output or no social copy. Use:

- 30 distinct title candidates by default, with one clearly marked as the recommended title; each title must be 15–20 visible characters (count Chinese characters, letters, digits, punctuation — every visible glyph counts). Titles shorter than 15 or longer than 20 visible characters are invalid and must be rewritten.
- A concise body that starts from the visual reading path, for example "通过这张图可以看到...".
- A mandatory separate `互动提问` section after the title candidates and before the platform rewrites.
- Every `互动提问` section must contain exactly 10 questions.
- Ask only knowledge- and content-related questions about the current subject.
- Focus on structure, component function, architecture, process, trade-offs, bottlenecks, applications, limitations, reliability, cost, efficiency, or technical comparisons.
- Phrase every item as a genuine question ending in `？` or `?`, for example asking what function a component performs, where a bottleneck occurs, why a material was selected, or what trade-off an architecture makes.
- Every question must explicitly include the topic keyword (the subject of the poster, e.g. NVLink, B200, CoWoS) in the question text. Do not use generic references like "this architecture", "this component", or "this design" as a substitute for the keyword.
- Do not ask what topic, product, teardown, article, image, or subject the audience wants next.
- Do not ask subjective-preference or opinion questions. Reject phrasing such as "你觉得", "你认为", "哪个更", "会更", "会不会", "值不值得", "有没有必要", or any question that asks the reader to pick a favorite, predict adoption, or express personal taste. Every question must have a verifiable technical answer grounded in the subject's structure, process, physics, or engineering constraints.
- Do not use generic engagement bait without a concrete technical object or trade-off.
- Reject prompts such as asking what to cover next, what else the reader wants to see, or requesting future-topic suggestions in comments.
- Do not write analytical evaluation paragraphs, generic praise, rhetorical promotion, predetermined conclusions, or unrelated engagement bait.
- Default platform rewrites for 作者, 小红书, 朋友圈/社群, and 知乎/B站动态.
- Append `声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。` in full directly after the content of each platform rewrite.
- Repeat the disclaimer for every platform even when the wording is identical. Never use one shared "通用结尾", reusable ending, or cross-reference in place of the repeated text.
- Keep platform rewrites concise, distinct, and non-promotional; avoid hype words, sales CTAs, and exaggerated claims.
- A request for comments or a full comment pack still receives exactly 10 knowledge-focused questions.
- For architecture diagrams, keep the body concrete: whole package, stacked layers, key interconnect, base/control layer, package/interposer, and what the structure enables.

Text-Risk Note

- Include a short risk note for prompt-only Chinese poster tasks when useful.
- Flag only obvious platform-risk terms: politics, sexual/adult content, graphic violence, gambling, drugs, hate/extremism, illegal activity, and similar categories.
- Treat technical false positives in context. `爆炸图/爆炸展开` are exploded-view terms; `安全隔离` is a security/networking term; `确定性` is a performance/architecture term.
- Do not claim visible image text has been inspected, because this prompt-only skill does not generate or verify images.

web search:
web search: Samsung HBM-PIM processing in memory official
codex
假设：按“概念爆炸图/示意结构图”处理，不刻画未公开版图；相关公司仅作为产业链注释随组件出现。

text
画布比例必须为 3:4 竖版（Portrait 3:4）。
不允许生成 2:3、4:5、9:16 或其他比例。
所有内容需完整落在安全区域内，约 5% 边距，标题、编号、标签、声明不得贴边或被裁切。
竖版 3:4 中文技术拆解海报，简体中文，现代无衬线字体，深色半导体实验室风格，精密芯片封装渲染 + 工程信息图排版。

主题：PIM 存内计算芯片，3D DRAM 计算近存储，示意结构图。
主标题：把计算塞进 3D DRAM
副标题：PIM 存内计算芯片如何减少数据搬运
画面中央放置完整芯片封装，沿垂直轴向上展开，形成“完整封装 -> 分层展开 -> 功能流向”的阅读路径。中间是高精度 3D DRAM 堆叠结构，左右两侧用细线 callout 标注功能，底部用小型流程条表现“数据驻留、近侧运算、结果回传”。整体科技感、清晰、克制，不要营销风。

编号组件层：
01 "完整芯片封装" - 顶部保护盖、封装体、低反射黑色基材，表现完整 PIM 存内计算芯片的外观 | 相关公司：长电科技、通富微电、华天科技
02 "DRAM 存储阵列层" - 多层 3D DRAM 堆叠，显示存储单元阵列、bank group、片内局部数据路径 | 相关公司：三星电子、SK hynix、Micron Technology
03 "近存储计算单元" - 在存储阵列附近放置轻量 MAC / SIMD 单元，表现向量乘加、激活、归一化等近侧计算 | 相关公司：寒武纪、景嘉微、安路科技
04 "TSV 垂直互连" - 贯穿堆叠的高密度垂直通道，用发光细柱表示高带宽、低延迟连接 | 相关公司：长电科技、通富微电、华天科技
05 "逻辑基底层" - 底部 logic base die，包含命令译码、调度、地址映射、数据路由 | 相关公司：澜起科技、芯原股份、聚辰股份
06 "片上网络与仲裁" - 细网格状 NoC，连接计算单元、DRAM bank、外部接口，表现并行访问与冲突仲裁 | 相关公司：芯原股份、寒武纪、安路科技
07 "电源时钟管理" - 小型 PMIC、PLL、稳压岛、时钟树，突出低功耗与稳定供电 | 相关公司：圣邦股份、思瑞浦、纳芯微
08 "热扩散与封装材料" - 热扩散片、底部填充、低翘曲材料，表现堆叠芯片的散热路径 | 相关公司：飞凯材料、鼎龙股份、彤程新材
09 "测试与良率监测" - 内建自测、冗余修复、探针测试标记，突出堆叠良率与可靠性 | 相关公司：华峰测控、精测电子、长川科技
10 "外部接口与基板" - 高密度封装基板、微凸点、外部高速接口，连接 AI 加速器或主处理器 | 相关公司：生益科技、兴森科技、胜宏科技

左侧信息面板：
"为什么靠近存储计算"
"减少 DRAM 与处理器之间的数据搬运，降低带宽压力与能耗。"

右侧信息面板：
"适合什么负载"
"矩阵运算、推荐系统、图计算、注意力相关的内存密集型任务。"

底部小流程条：
"数据留在阵列附近 -> 近侧执行简单计算 -> 压缩结果回传"
"核心不是替代 GPU，而是缓解 memory wall。"

图中必须包含来源声明，放在下方中部的工程注释区，清晰可读但小于主标签：
"声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。"

视觉要求：
高精度半导体封装质感，透明分层，微凸点、TSV、硅基底、封装基板清晰可见；编号圆点统一；公司名称只作为对应组件行末的小字注释，样式完全一致；所有文字为简体中文或必要英文缩写，如 PIM、DRAM、TSV、NoC、MAC、SIMD；避免乱码、日文、伪文字、重复文字、过密小字、标签重叠、部件漂浮无连接。

禁止：
不要出现任何官方 logo、官方标识、官方字形、商标图形、企业徽记；不要生成公司图标；不要写公司介绍、产品名、股票代码、链接、排名、所在地；不要使用夸张营销词；不要把声明做成水印；不要把声明放到画布边缘；不要画成真实未公开芯片版图。

发布文案

标题候选
1. 把计算塞进3D DRAM
2. PIM芯片为何靠近存储
3. 3D DRAM里的计算层
4. 存内计算芯片结构拆解
5. 近存储计算怎么省带宽
6. PIM如何缓解内存墙
7. DRAM堆叠里的计算逻辑
8. 存内计算不是普通内存
9. 3D DRAM近侧运算路径
10. PIM芯片的分层秘密
11. 数据不出阵列也能算
12. 计算靠近DRAM的理由
13. 一图看懂PIM芯片结构
14. 3D存储堆叠如何计算
15. PIM架构拆开看明白
16. 存内计算的带宽答案
17. DRAM旁边为什么放计算
18. PIM芯片不是简单堆叠
19. 近存储计算的关键部件
20. 从TSV看懂PIM结构
21. PIM如何减少数据搬运
22. 存储芯片里的计算通道
23. 3D DRAM计算层示意图
24. PIM芯片封装分层解析
25. 内存墙背后的PIM思路
26. 近存储计算芯片怎么拆
27. PIM里的逻辑基底层
28. DRAM计算近存储图解
29. PIM芯片的互连与调度
30. 3D DRAM为何适合PIM

互动提问
1. PIM 存内计算芯片中的 TSV 主要解决哪类带宽瓶颈？
2. PIM 存内计算芯片为何需要逻辑基底层承担调度？
3. PIM 存内计算芯片的近存储计算单元通常适合哪些算子？
4. PIM 存内计算芯片如何减少 DRAM 与处理器之间的数据搬运？
5. PIM 存内计算芯片的热扩散路径会受哪些堆叠因素影响？
6. PIM 存内计算芯片中的片上网络如何处理并行访问冲突？
7. PIM 存内计算芯片的内建测试为何会影响堆叠良率？
8. PIM 存内计算芯片与传统 HBM 的结构差异主要在哪里？
9. PIM 存内计算芯片的电源时钟管理需要控制哪些噪声来源？
10. PIM 存内计算芯片在 AI 负载中通常缓解哪一类内存墙问题？

版本一
通过这张图可以看到，PIM 存内计算芯片的关键不是“把芯片堆高”，而是把轻量计算单元放到 3D DRAM 阵列附近。数据少走一段路，带宽、延迟、能耗都会被重新分配。

图里重点拆了 DRAM 阵列层、近存储计算单元、TSV 垂直互连、逻辑基底层、片上网络、电源时钟、热扩散与测试良率。它更像是在内存旁边加一个可控的计算邻居，而不是替代主处理器。
声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。

版本二
PIM 存内计算芯片解决的是一个很朴素的问题：大量数据在处理器和内存之间来回搬，能耗和带宽都被消耗在路上。

3D DRAM 结构给了它新的空间：存储阵列负责保存数据，近侧计算单元处理适合的简单算子，TSV 和逻辑基底层负责把访问、调度、回传组织起来。看懂这张分层图，就能理解 PIM 为什么总和内存墙一起被讨论。
声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。

版本三
这张图按“完整封装 -> 分层展开 -> 数据路径”的顺序拆 PIM 存内计算芯片。

最核心的结构是 3D DRAM 堆叠、近存储计算单元、TSV 垂直互连和底部逻辑层。它的价值不在于做所有计算，而是在内存密集型任务里，把部分可并行、可局部处理的计算放到数据附近完成。
声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。

版本四
如果把传统计算看成“数据去找处理器”，PIM 存内计算芯片更像是“让一部分计算去找数据”。

3D DRAM 的堆叠结构提供高带宽通道，近存储计算单元处理局部计算，逻辑基底层负责调度和接口，封装材料与测试结构则决定它能不能稳定制造出来。真正的难点不只在架构，也在封装、互连、功耗和良率。
声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。

未生成图片，也未做视觉检查。技术参考：Samsung HBM-PIM 官方介绍、SK hynix GDDR6-AiM 资料、KAIST 3D IC near-memory computing 页面。
tokens used
37,265
假设：按“概念爆炸图/示意结构图”处理，不刻画未公开版图；相关公司仅作为产业链注释随组件出现。

text
画布比例必须为 3:4 竖版（Portrait 3:4）。
不允许生成 2:3、4:5、9:16 或其他比例。
所有内容需完整落在安全区域内，约 5% 边距，标题、编号、标签、声明不得贴边或被裁切。
竖版 3:4 中文技术拆解海报，简体中文，现代无衬线字体，深色半导体实验室风格，精密芯片封装渲染 + 工程信息图排版。

主题：PIM 存内计算芯片，3D DRAM 计算近存储，示意结构图。
主标题：把计算塞进 3D DRAM
副标题：PIM 存内计算芯片如何减少数据搬运
画面中央放置完整芯片封装，沿垂直轴向上展开，形成“完整封装 -> 分层展开 -> 功能流向”的阅读路径。中间是高精度 3D DRAM 堆叠结构，左右两侧用细线 callout 标注功能，底部用小型流程条表现“数据驻留、近侧运算、结果回传”。整体科技感、清晰、克制，不要营销风。

编号组件层：
01 "完整芯片封装" - 顶部保护盖、封装体、低反射黑色基材，表现完整 PIM 存内计算芯片的外观 | 相关公司：长电科技、通富微电、华天科技
02 "DRAM 存储阵列层" - 多层 3D DRAM 堆叠，显示存储单元阵列、bank group、片内局部数据路径 | 相关公司：三星电子、SK hynix、Micron Technology
03 "近存储计算单元" - 在存储阵列附近放置轻量 MAC / SIMD 单元，表现向量乘加、激活、归一化等近侧计算 | 相关公司：寒武纪、景嘉微、安路科技
04 "TSV 垂直互连" - 贯穿堆叠的高密度垂直通道，用发光细柱表示高带宽、低延迟连接 | 相关公司：长电科技、通富微电、华天科技
05 "逻辑基底层" - 底部 logic base die，包含命令译码、调度、地址映射、数据路由 | 相关公司：澜起科技、芯原股份、聚辰股份
06 "片上网络与仲裁" - 细网格状 NoC，连接计算单元、DRAM bank、外部接口，表现并行访问与冲突仲裁 | 相关公司：芯原股份、寒武纪、安路科技
07 "电源时钟管理" - 小型 PMIC、PLL、稳压岛、时钟树，突出低功耗与稳定供电 | 相关公司：圣邦股份、思瑞浦、纳芯微
08 "热扩散与封装材料" - 热扩散片、底部填充、低翘曲材料，表现堆叠芯片的散热路径 | 相关公司：飞凯材料、鼎龙股份、彤程新材
09 "测试与良率监测" - 内建自测、冗余修复、探针测试标记，突出堆叠良率与可靠性 | 相关公司：华峰测控、精测电子、长川科技
10 "外部接口与基板" - 高密度封装基板、微凸点、外部高速接口，连接 AI 加速器或主处理器 | 相关公司：生益科技、兴森科技、胜宏科技

左侧信息面板：
"为什么靠近存储计算"
"减少 DRAM 与处理器之间的数据搬运，降低带宽压力与能耗。"

右侧信息面板：
"适合什么负载"
"矩阵运算、推荐系统、图计算、注意力相关的内存密集型任务。"

底部小流程条：
"数据留在阵列附近 -> 近侧执行简单计算 -> 压缩结果回传"
"核心不是替代 GPU，而是缓解 memory wall。"

图中必须包含来源声明，放在下方中部的工程注释区，清晰可读但小于主标签：
"声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。"

视觉要求：
高精度半导体封装质感，透明分层，微凸点、TSV、硅基底、封装基板清晰可见；编号圆点统一；公司名称只作为对应组件行末的小字注释，样式完全一致；所有文字为简体中文或必要英文缩写，如 PIM、DRAM、TSV、NoC、MAC、SIMD；避免乱码、日文、伪文字、重复文字、过密小字、标签重叠、部件漂浮无连接。

禁止：
不要出现任何官方 logo、官方标识、官方字形、商标图形、企业徽记；不要生成公司图标；不要写公司介绍、产品名、股票代码、链接、排名、所在地；不要使用夸张营销词；不要把声明做成水印；不要把声明放到画布边缘；不要画成真实未公开芯片版图。

发布文案

标题候选
1. 把计算塞进3D DRAM
2. PIM芯片为何靠近存储
3. 3D DRAM里的计算层
4. 存内计算芯片结构拆解
5. 近存储计算怎么省带宽
6. PIM如何缓解内存墙
7. DRAM堆叠里的计算逻辑
8. 存内计算不是普通内存
9. 3D DRAM近侧运算路径
10. PIM芯片的分层秘密
11. 数据不出阵列也能算
12. 计算靠近DRAM的理由
13. 一图看懂PIM芯片结构
14. 3D存储堆叠如何计算
15. PIM架构拆开看明白
16. 存内计算的带宽答案
17. DRAM旁边为什么放计算
18. PIM芯片不是简单堆叠
19. 近存储计算的关键部件
20. 从TSV看懂PIM结构
21. PIM如何减少数据搬运
22. 存储芯片里的计算通道
23. 3D DRAM计算层示意图
24. PIM芯片封装分层解析
25. 内存墙背后的PIM思路
26. 近存储计算芯片怎么拆
27. PIM里的逻辑基底层
28. DRAM计算近存储图解
29. PIM芯片的互连与调度
30. 3D DRAM为何适合PIM

互动提问
1. PIM 存内计算芯片中的 TSV 主要解决哪类带宽瓶颈？
2. PIM 存内计算芯片为何需要逻辑基底层承担调度？
3. PIM 存内计算芯片的近存储计算单元通常适合哪些算子？
4. PIM 存内计算芯片如何减少 DRAM 与处理器之间的数据搬运？
5. PIM 存内计算芯片的热扩散路径会受哪些堆叠因素影响？
6. PIM 存内计算芯片中的片上网络如何处理并行访问冲突？
7. PIM 存内计算芯片的内建测试为何会影响堆叠良率？
8. PIM 存内计算芯片与传统 HBM 的结构差异主要在哪里？
9. PIM 存内计算芯片的电源时钟管理需要控制哪些噪声来源？
10. PIM 存内计算芯片在 AI 负载中通常缓解哪一类内存墙问题？

版本一
通过这张图可以看到，PIM 存内计算芯片的关键不是“把芯片堆高”，而是把轻量计算单元放到 3D DRAM 阵列附近。数据少走一段路，带宽、延迟、能耗都会被重新分配。

图里重点拆了 DRAM 阵列层、近存储计算单元、TSV 垂直互连、逻辑基底层、片上网络、电源时钟、热扩散与测试良率。它更像是在内存旁边加一个可控的计算邻居，而不是替代主处理器。
声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。

版本二
PIM 存内计算芯片解决的是一个很朴素的问题：大量数据在处理器和内存之间来回搬，能耗和带宽都被消耗在路上。

3D DRAM 结构给了它新的空间：存储阵列负责保存数据，近侧计算单元处理适合的简单算子，TSV 和逻辑基底层负责把访问、调度、回传组织起来。看懂这张分层图，就能理解 PIM 为什么总和内存墙一起被讨论。
声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。

版本三
这张图按“完整封装 -> 分层展开 -> 数据路径”的顺序拆 PIM 存内计算芯片。

最核心的结构是 3D DRAM 堆叠、近存储计算单元、TSV 垂直互连和底部逻辑层。它的价值不在于做所有计算，而是在内存密集型任务里，把部分可并行、可局部处理的计算放到数据附近完成。
声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。

版本四
如果把传统计算看成“数据去找处理器”，PIM 存内计算芯片更像是“让一部分计算去找数据”。

3D DRAM 的堆叠结构提供高带宽通道，近存储计算单元处理局部计算，逻辑基底层负责调度和接口，封装材料与测试结构则决定它能不能稳定制造出来。真正的难点不只在架构，也在封装、互连、功耗和良率。
声明：信息整理自网络公开内容，仅供学习交流；如有疏漏或错误，请联系指正。

未生成图片，也未做视觉检查。技术参考：Samsung HBM-PIM 官方介绍、SK hynix GDDR6-AiM 资料、KAIST 3D IC near-memory computing 页面。