The mlsysim docs hero linked "Slide Decks" to mlsysbook.ai/slides/, the
textbook teaching site, which has no relationship to the MLSys·im ISCA
tutorial. Wire the tutorial decks into the same publish path the paper
uses so the link resolves to mlsysim's own slides.
- mlsysim-build-pdfs.yml: split into build-paper / build-slides; new
job installs xelatex + JetBrains Mono and uploads MLSYSIM-Slides
- mlsysim-publish-live.yml, mlsysim-preview-dev.yml: download
MLSYSIM-Slides and inject tutorial_part{1,2}.pdf into MLSYSIM_DOCS
- mlsysim-update-pdfs.yml: redeploy slide PDFs in PDF-only hot-fix path
- mlsysim/tutorial/Makefile: build both decks (was part1 only) with
proper target tracking
- mlsysim/docs/slides.qmd: new landing page with download buttons and
a pointer disambiguating the textbook lecture decks
- mlsysim/docs/index.qmd: hero CTA now targets slides.qmd
- mlsysim/docs/config/_quarto-html.yml: register slide PDFs as
resources and add navbar entry
Align the MLSys·im code, docs, paper, website, workflows, and lab wheel for the 0.1.1 release. This also fixes runtime/API issues found during release review and prepares the paper PDF plus archive package.
Four user-facing identity statements referenced an earlier working title
("A Composable Analytical Framework for Machine Learning Systems") that
no longer matches the actual paper title fixed in 0.1.1 ("MLSys·im:
First-Principles Infrastructure Modeling for Machine Learning Systems").
Align each identity-claiming statement to the paper title.
This covers user-facing *name-claims* only — the places where mlsysim
describes itself. Descriptive uses of "analytical framework" as a
technical category inside the paper and related technical prose are
retained (they situate mlsysim among other analytical tools like
Paleo, Calculon, Vidur; those uses are legitimate).
- mlsysim/pyproject.toml : project description
- mlsysim/mlsysim/cli/main.py : `mlsysim --help` text
- mlsysim/docs/tutorials/index.qmd: tutorial landing blurb
- mlsysim/tutorial/prerequisites.md: prerequisites preamble
Replaced bare Metropolis with the mlsys Beamer theme used by the
textbook lecture slides. Both decks now have:
- Crimson section navigation bar in header
- Branded footer (book logo + Harvard logo + page numbers)
- Helvetica Neue body font + JetBrains Mono/Courier code font
- Section divider pages with progress indicator
- \mlsysfocus for key equation slides
Part 1: 90 pages, 0 errors. Part 2: 47 pages, 0 errors.
Eta explanation, AllReduce numbers-first, and compression fleet economics
all validated by re-simulated personas. 2 of 5 issues fully resolved,
2 partially resolved (timing still tight), 1 unaddressed (TinyML depth).
Multi-vendor Roofline is the single biggest NPS driver.
1. Added 'Understanding eta' slide with table and CPI analogy (before first demo)
2. Added concrete AllReduce example with numbers BEFORE formula
3. Added 'Compression Changes Fleet Architecture' slide with cost table
Addresses top 3 issues from simulation feedback round 1.
- Changed A100-vs-H100 comparisons to H100-vs-MI300X-vs-Gaudi3 three-way
- Added Hardware Zoo slide showing all 22 platforms across 6 vendors
- Added timing markers [N min] to speaker notes throughout
- Added "turn to your neighbor" discussion prompts after predict slides
- Added roadmap "You Are Here" slides after each break
- Stripped code boilerplate (import only shown once, aliases for later demos)
- Added figure references for new SVGs (multi-vendor-roofline, parallelism-3d,
memory-hierarchy, pareto-front)
- Updated related_work.tex comparison table to highlight vendor breadth
1. Added $200M opening hook (2 slides before learning objectives)
2. Wired \input{related_work} between Part 0 and Part 1
3. Replaced figure placeholders with real SVG PDFs (carbon, hardware)
4. Added predict-then-reveal prompts before Aha #3 (256 GPU scaling)
and Aha #4 (INT4 fleet halving)
5. Added "Name That Wall" personal transfer moment before closing
Addresses all 5 P0 blockers from Patterson's presentation review.
All font-size values < 10px bumped to 9.5-11px minimum across all 6
figures. At 900x500 viewBox on a 16:9 projection, this ensures text
is readable from the back of a 300-person room.
Covers: install, first demo, first homework, semester plan, auto-grading
hints, and material inventory. Designed to minimize instructor adoption
friction for ML systems courses.
Parts 5-9 were lecture-only (0 code blocks). Now every section has
live demo code using real mlsysim API: EconomicsModel, SustainabilityModel,
DSE search, Engine.sweep, ESP32/nRF52840/H100 comparison, Pipeline
composition, SensitivitySolver, capstone starter code.
Figure placeholders added for carbon-geography and hardware-spectrum SVGs.
Supports inkscape and rsvg-convert for SVG conversion.
Targets: make (all), make svgs, make slides, make check, make clean.
Build output goes to _build/ (gitignored).
Within-subjects design (N≥30), paired t-test analysis plan, IRB
considerations, expected effect size (d=0.8), and timeline for
running at ISCA 2026. Produces publishable data for SIGCSE/L@S.
Comparison table of 7 tools (ASTRA-sim, Calculon, DeepSpeed, LLMPerf, etc.).
Fidelity-speed spectrum TikZ diagram. Honest 'What We Don't Do' slide.
Iron Law wall-to-term mapping. CPI analogy for the efficiency parameter.
22 walls at a glance compact table. Designed for \input{} into main deck.
Designed for dual use: tutorial engagement + publishable research data.
Tests transfer (apply framework to unfamiliar systems), not recall.
Includes distractor analysis and scoring rubric.
