Vijay Janapa Reddi
6cc94912eb
feat(staffml): 9,126 Qs, 87 topics, 0 ghost topics, sustainability added
...
All 87 topics now populated (was 5 ghost topics). New topics:
autograd (50), chiplet (25), model-adaptation (25), disaggregated-
serving (25), recommendation-systems (25), sustainability (50),
software-portability (50), comm-compute-overlap (50).
RAI expanded (differential-privacy, fairness, responsible-ai).
Vendor diversity improved (MI300X/TPU 30%+ in new cloud Qs).
Phase metadata reclassified. Paper macros auto-updated.
Figures rebuilt. 19/19 invariants pass. 88.3% validated.
2026-04-02 15:15:11 -04:00
Vijay Janapa Reddi
e343b9f589
docs(staffml): mock NeurIPS D&B reviews + paper fixes + missing references
...
4 mock NeurIPS reviews: avg Novelty 5.8, Quality 4.8, Significance 6.3,
Clarity 7.0. Universal feedback: pilot study needed for D&B acceptance.
Paper fixes applied:
- Bloom critique softened (complements, not departs from)
- LLM generation ratio disclosed (95% + 4.2% error rate)
- Scope explicitly limited to technical systems reasoning
- Added MMLU, HumanEval, AERA testing standards references
- Internal consistency fixes (macros, track %, caption numbers)
2026-04-02 07:40:54 -04:00
Vijay Janapa Reddi
9955a76b92
feat(staffml): deep verification + mock NeurIPS reviews + paper improvements
...
Deep verification: 237-question stratified sample, 4.2% error rate found.
All 10 errors fixed (unit confusion, arithmetic, conceptual misapplication).
96 physics violations removed (impossible topic×track pairs).
Extended invariant checks added (applicability matrix enforcement).
Paper improvements from mock NeurIPS review feedback:
- Bloom critique softened ("complements" not "departs from")
- LLM generation transparency (95% ratio + 4.2% error rate disclosed)
- Scope explicitly limited to technical systems reasoning
- H100 specs corrected (989 TFLOPS, not 495)
- Track percentages reference table instead of hardcoding
- Figure captions use macros for consistency
New topics with questions: software-portability (50), comm-compute-overlap (50).
Phase metadata reclassified (42.5% inference, 37.7% both, 19.9% training).
2026-04-02 07:28:41 -04:00
Vijay Janapa Reddi
0488d01c79
feat(tutorial): add 38 code listings + figure refs to Part 2 slides
...
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.
2026-04-01 19:23:31 -04:00
Vijay Janapa Reddi
42dc31a202
feat(tutorial): 6 publication-quality SVG figures for ISCA slides
...
- roofline-model.svg: Classic Roofline with LLM decode + CNN training points
- iron-law-decomposition.svg: Iron Law equation with wall-to-term mapping
- serving-two-phases.svg: Prefill (compute) vs Decode (memory) phases
- allreduce-ring.svg: 8-GPU ring with reduce-scatter + all-gather
- hardware-spectrum.svg: nRF52840 → ESP32 → Jetson → H100 → NVL72 scale
- carbon-geography.svg: Norway/Quebec/US/Poland bar chart (41x gap)
All follow svg-style.md: 900x500 viewBox, semantic colors, Helvetica font.
2026-04-01 19:22:20 -04:00
