Skip to content

PLAN.md

Latest commit

 

History

History
115 lines (90 loc) · 6.63 KB

PLAN.md

File metadata and controls

115 lines (90 loc) · 6.63 KB
Raw

Medium starter plan — Parallel Spectrum

How the curriculum becomes a Medium publication. Source packs live in ../social/ (feed-length) and ../A0x-*/ (video studio); this plan is the long-form text channel of the same measured stories.

Brand: ../../docs/BRAND.md · Social chain: ../social/HISTORY.md

1. Positioning

Publication name: Parallel Spectrum Tagline: Measured GPU engineering — kernels, bandwidth, and the honest path between them. Angle that differentiates: every post leads with a real profiled number and ends with an honest assessment (what broke, when not to use it, what to verify). No hype, no vanity benchmarks.

One publication (not scattered profile posts) so the series accumulates followers and a table of contents.

2. Setup checklist (one-time)

  • Create Medium publication named Parallel Spectrum (not just profile posts)
  • Avatar/header: slate #0F172A + cyan #0891B2 bar (brand, no rainbow, no vendor logos)
  • Homepage layout: one featured story (post 001) + a "Series" section per track
  • About page: 3 lines — what you measure, who it's for, link to GitHub repo
  • Set publication SEO description: "GPU programming and ML performance, taught with real profiles — HIP, CUDA, Triton"
  • Connect social accounts; enable newsletter on the publication

3. Editorial calendar — first 8 posts

Cadence: 1 post/week. LinkedIn + Short echo within 48 h of each Medium post (same hook number).

Wk Social # Module Working title (names the outcome, not the topic) Hook number Source status
1 001 A01 Your GPU kernel took 44 µs. The program took 72 ms. kernel = 0.06% of wall clock ✅ blog.md ready — re-run first
2 002 A02 Your GPU kernel passed. It still wasted ~85% of bandwidth ~700 → ~3,400 GB/s ✅ content/social/002-a02-coalescing-bandwidth/blog.md ready — re-run first
3 003 A01 This kernel corrupts memory and still prints the right answer N=1000, block=256 → 24 threads OOB ✅ content/social/003-a01-bounds-silent-corruption/blog.md ready — arithmetic, no GPU re-run
4 004 A04 Nine ways to reduce: the reduction ladder's stage-1-to-stage-9 spread TBD — re-run ladder ⬜ needs measurement
5 005 A06 Naive vs tiled GEMM: what shared memory actually buys TBD — re-run GEMM ⬜ needs measurement
6 006 A07 The poisoned row: one nan and your softmax lies poisoned-row repro ⬜ needs measurement
7 007 B02 The roofline knee your sweep finds by itself ridge-point crossover ⬜ needs measurement
8 008 B03 Profile first, optimize second: a bottleneck triage walkthrough TBD ⬜ needs measurement

Hard rule from the content studio: re-run the lab on your own GPU before publishing; quote your numbers and name the GPU. Unattributed numbers are worse than none.

Track mix: first 3 are Track A (cyan #0891B2). From week 4 alternate tracks so the publication doesn't read as only-kernels: A → A → A → A → A → A → B → B → C (C01 Problem framing is the natural week-9 palate cleanser).

4. Medium post template (per post)

Source = content/social/NNN-<slug>/blog.md expanded to ~800–1,500 words (7-min read). The social stub is the skeleton; add the module's Section 2 first-principles excerpt in the middle.

Title       — outcome + number ("...44 µs... 72 ms"), not a topic label
Subtitle    — one line: what the reader will be able to assess after
Hero image  — Parallel Spectrum visual from the pack (visual.png); alt-text required
Kicker      — "Parallel Spectrum · Post NNN · Track X"
Sections:
  1. TL;DR (3 sentences)
  2. What we measured — phase/bandwidth table
  3. Why (first principles, excerpted from module README S2)
  4. The snippet (short fenced block; gist embed if >30 lines)
  5. Reproduce (one command block)
  6. Honest assessment — watch for / when not to / I'd verify
  7. Series links — prev/next post + module folder + repo

Medium-specific mechanics

  • Mermaid does not render on Medium. Export diagrams as PNG (brand init block intact) and embed as images with alt-text. GitHub versions keep live Mermaid.
  • Code: fenced blocks with language tag for ≤30 lines; GitHub gist embeds for longer so readers can fork. Never screenshots of code.
  • Tags (max 5): pick from GPU Programming CUDA ROCm Parallel Computing Machine Learning Performance Engineering — keep 2 stable across the series for follow-through.
  • Canonical: if a version first appears on a personal blog/GitHub Pages, use Medium's Import story tool (medium.com/p/import) so canonical points to the original. If Medium is primary, publish native and link out — no import needed.
  • SEO: set custom SEO title/description in story settings; first paragraph must work as the search snippet (that's where the hook number goes).
  • End card: fixed footer — "Parallel Spectrum · Post NNN · Prev: … · Next: … · Code: …"

5. Cross-channel loop (each post)

Medium (long-form text, week N)
  ├─▶ LinkedIn post (≤48 h later) — hook number + "full write-up" link
  ├─▶ Short/Reel (same week) — 55 s cut from content/<module>/shorts.md
  └─▶ GitHub — module folder already linked; add Medium URL to social pack blog.md

Then: update ../social/HISTORY.md — Publish log row per channel with real URL and date. The chain is the single source of "what went out when."

6. Publish-day checklist

  • Lab re-run on my GPU; numbers in post match my terminal
  • GPU + ROCm/CUDA version named in the post
  • Hero image alt-text filled; Mermaid exported as PNG
  • Title names the outcome; subtitle names the reader's takeaway
  • Kicker + footer series links correct (prev/next match HISTORY)
  • 5 tags set; SEO title/description set
  • Code runs from the reproduce block verbatim
  • After publish: HISTORY publish log + LinkedIn/Short echo scheduled

7. Immediate next actions

  1. Re-run A01 (make run-hip + make profile-hip), confirm/adjust the 44 µs / 72 ms figures in ../social/001-a01-kernel-vs-path/blog.md
  2. Expand that blog.md to ~1,000 words using module A01 Section 2 (first principles)
  3. Create the publication (Section 2 checklist) and import/publish post 001
  4. Log the Medium URL in HISTORY; then commission pack 002 (A02) with linkedin-shorts-maker