# Foundations figure sources

Reviewed 2026-10-13T05:00:08.587Z. Complete source PDFs remain temporary study copies outside the repository and deployment.

## Calculated memory budget

- Asset: `memory-budget.svg`, SHA-256 `10ee0fba06901c82a646f0fee780c9f8867341c7445cd36f4729eb578cc3a9bf`, viewBox 1000 by 466.
- Independent Latent Light SVG, generated by `docs/research/edge-ai/foundations/visuals/memory-budget.mjs`. The executable generator retains the exact assumptions and calculations; no source artwork is copied.
- Inputs: 3 billion stored parameters, four bits per weight, 28 cached layers, batch one, equal key/value head width 128, 16 bits per cache value. The three cases use 4,096/8, 8,192/8 and 4,096/2 token positions/KV heads.
- Ideal weights: 1.5 billion bytes, about 1.40 GiB. Ideal KV: 448, 896 and 112 MiB. Weight-plus-cache totals: 1.8345, 2.2720 and 1.5064 GiB, rounded to 1.83, 2.27 and 1.51 GiB.
- Mechanism source: [GQA v3](https://arxiv.org/html/2305.13245v3), Introduction and Section 2.2. The arithmetic is our derivation, not a reported experiment. Scales, metadata, live tensors, runtime workspace and application memory are excluded; the chart does not predict a process peak, task quality or latency.
- Original diagram and generator use the repository MIT license.

## Historical Roofline envelope

- Authors: Samuel Williams, Andrew Waterman and David Patterson.
- Source: [Roofline: An Insightful Visual Performance Model for Floating-Point Programs and Multicore Architectures](https://www2.eecs.berkeley.edu/Pubs/TechRpts/2008/EECS-2008-134.html), UCB/EECS-2008-134, October 17, 2008; Section 3, original Figure 1, PDF page 4, printed page 2.
- Exact PDF: https://www2.eecs.berkeley.edu/Pubs/TechRpts/2008/Archive/EECS-2008-134.pdf
- Original PDF SHA-256: `c92f852eed1070b140302114339271dc1a4d6d665814ac9000be6ce06a7d6bcb`; 16 pages, 612 by 792 PDF points. Temporary study copy: `latent-edge-roofline-2008.pdf`.
- Asset: `roofline-original-figure1.png`, SHA-256 `8f00c6df7380022158a0f014b3af2393f8d0cbc262428abc7e7de1d2e5e85d62`, 1035 by 680 pixels.
- Extraction: Poppler pdftoppm, PDF page 4 at 300 dpi, top-left crop x=1318, y=292, width=1035, height=680 pixels; equivalently x=316.32, y=70.08, width=248.4, height=163.2 PDF points. Both complete chart panels, axes, labels and original colors are retained. The surrounding prose caption is omitted; no data, annotation or plot is redrawn.
- The original figure compares historical Opteron X2 and X4 compute/bandwidth envelopes. It is not a phone, GPU or NPU benchmark. The article directly reads and qualifies the left memory/compute limit and the right higher-compute comparison.

### Specific criticism and review assessment

Copyright © 2008 by the authors; all rights reserved. The inspected PDF notice grants bounded personal/classroom noncommercial copying and separately restricts server posting. That notice is **not** asserted to authorize public site or repository reuse. No permissive public artwork license or direct author permission was established.

The assessed editorial basis is criticism and review under [Canada Copyright Act section 29.1](https://laws-lois.justice.gc.ca/eng/acts/c-42/Section-29.1.html), including source and author attribution, and the contextual fairness framework in the site's Content standards. This is not legal advice, a judicial determination or a worldwide reuse license.

- Purpose: analyze the authors' original performance envelope, its two panels and the limitation that a higher compute peak does not improve every memory-limited workload. The article connects it to a qualified time-bound derivation rather than treating it as generic decoration.
- Character and distribution: a public educational notebook, public GitHub source and standalone asset URL are explicitly considered. Attribution does not by itself grant rights.
- Amount: one two-panel technical figure, without the prose caption, nearby paper text, later figures or the complete paper. A complete figure can itself be a protected work; proportionality is considered in this specific explanatory context, not inferred from the paper's page count.
- Alternatives: an independent Roofline diagram would teach the generic relationship, but would not support this close reading of the actual historical X2/X4 comparison. The page already uses original synthesis for its memory and lifecycle explanations.
- Nature and effect: published scientific analysis made publicly available by the university. This excerpt and commentary are not a substitute for the report, its derivation or performance experiments, and do not bypass a paywall or create an image collection for sale. No separate figure-licensing market was found in the inspected materials; that is not proof that none exists.
- Third-party material: selected panels contain plotted technical relationships and hardware identifiers, not photographs or separately credited artwork. No trademark grant or hardware endorsement is implied.
- Remaining uncertainty: fairness depends on facts and jurisdiction. Reassess changed distribution, commercial packaging or wider reuse; seek permission if needed.

Source edition, PDF page, complete panels and crop were visually inspected. A prior failed PDF access is retained in the research history and was resolved by this exact download. The original plot excerpt is excluded from Latent Light's MIT license.
