Paper VIII visual guide · Visual 26 of 32 · A BENCHMARK ROW IS A CONTRACT

Same output before comparing times

Traditional pp. 9, 36, 37 / Proposition 2.9, Section 11, Section 12.

Start reading

Resources & source

Original plate in the visual-guide PDF · Original atlas source

Skip contents and start reading

Consider four synthetic timing rows for a product. Row B returns factors, while row A returns an explicit matrix. A ratio \(12/2=6\) does not compare two implementations of the same returned object. Adding the decoding cost changes the candidate to row C.

Matching output contracts before comparing methods
Original drawing. Read the description below; expand the labels and data for selectable text.

Reading the drawing

Each route identifies the object it returns. A timing comparison requires the same requested output, validation target and measurement conditions.

Drawing labels and data 21 labels

Read the original labels as selectable text. Tables preserve grid relationships; other short labels are grouped compactly.

  • row
  • output
  • scope
  • time
  • A
  • explicit matrix
  • materialization included
  • 12
  • B
  • low-rank factors
  • factor product only
  • 2
  • C
  • explicit matrix
  • factor product + decode
  • 14
  • D
  • explicit matrix
  • different input precision
  • 1

Synthetic times: A versus C is contract-aligned. A versus B is not a sixfold speedup.

Open the full-size drawing

What a fair match retains

The input values, dimensions, layout, source precision, correctness reference, returned output, and timing scope must agree. Row D changes the input precision; that can be a useful accuracy tradeoff, but it is not the same case merely because the output has the same shape.

If a downstream task consumes factors directly, row B has a valid compressed-output contract. It should be compared with another method that supplies that downstream result, not forced to materialize an unused array. Fairness does not mean every algorithm must use the same internal representation.

How the original measurements are treated here

The preserved manuscript contains its historical CPU/GPU corpus. Replaying its figure generator redraws those saved observations. It does not rerun the recorded device, establish present hardware performance, or transform logical avoided work into physically executed floating-point operations. No new GPU timing is added by this guide.

Exact representation gains, device gains, kernel gains, and amortization gains can be isolated only through appropriately matched rows. A total ratio that mixes them is still an observation, but not a decomposition theorem.

Read this as

A timing number without its case and scope is incomplete data. Preserve the mathematical output contract before claiming a speedup.

Source and conversion record

The original atlas wording, equations, and drawing labels are retained. Drawings are views of the retained visual-guide PDF. Text descriptions and table headings are added reading aids. This conversion does not rebuild TeX, rerun the examples, audit proofs or compile Lean.

The revised crosswalk binds all 32 plates to the retained PDF. Manuscript links use stable source labels checked against the manuscript. The guide’s “Classical input 7.5” resolves to the manuscript’s “Standard fact 7.5” at the same label.

Conversion checks and source hashes

Notation in Paper VIII

Look up a symbol without leaving this chapter. Select text before opening help to search for it.

Search this work’s notation.

    Symbols can have different meanings in different passages. Read each result in its own context.

    Open the full notation index