micro-writing-style

MICRO Writing Style

A MICRO paper is an engineering argument: this pattern exists in real workloads, here is a piece of hardware that exploits it, here is what it costs, here is what it pays. The style rules below encode how the microarchitecture community reads.

The four-beat argument

1. Characterization beat. Open with measured evidence that the exploitable behavior exists — a miss-pattern histogram, a stall-cycle breakdown, a reuse-distance CDF — from real workloads, before any mechanism appears. MICRO reviewers trust a paper that observed the machine before modifying it.
2. Insight beat. One sentence naming the property your hardware will exploit ("dead blocks announce themselves by burst-then-silence access"). If the insight sentence could headline three different mechanisms, sharpen it.
3. Mechanism beat. A block diagram plus a walkthrough at the level of tables, counters, comparators, and state machines — detailed enough that a reader could size the SRAM. Include the storage budget table (bits per entry × entries).
4. Accounting beat. Speedup is half a result. The other half: area or storage overhead, power, added latency on the critical path, and what happens on the workloads the mechanism does not help (the tax on non-beneficiaries).

Page budget for 11 content pages, no appendix

The 2026 guidelines permit no appendix, so the body must be self-sufficient — proofs of concept, sensitivity sweeps, and methodology detail all compete for the same 11 pages (references are unlimited and free).

Squeezing tricks (\vspace, tightened lists, sub-9pt figure text) are barred by the guidelines and checked visually — recover space by cutting content, not leading.

Sentence-level calibration

• Tie every performance number to its instrument: "8.4% geomean speedup on SPEC CPU2017 rate-1 in cycle-level simulation" — never let a simulated number read as a silicon number.
• Prefer mechanism verbs (filters, throttles, bypasses, coalesces) over marketing verbs (unlocks, revolutionizes, dramatically improves).
• State the baseline aggressively and early: "against an 8-way OoO core with a state-of-the-art stride+SMS prefetcher" beats "against a baseline core."
• Quantify the negative space: which workload class regresses, by how much, and why the mechanism's guard logic bounds it.

Figure discipline

Fig. 1  motivating characterization (the "why" figure)   — page 1-2
Fig. 2  mechanism block diagram (the "what" figure)      — mechanism section
Fig. 3  headline bar chart, geomean marked, baseline = 1 — first evaluation page
Fig. 4+ ablation / sensitivity, one question per figure
Every figure: axis labels >= 9pt after scaling; caption states config + workload set.

A MICRO reader flips to Fig. 2 and Fig. 3 before reading a word — both must stand alone with their captions.

Naming and terminology conventions

• Name the mechanism once, early, and use the name relentlessly. MICRO papers are remembered by mechanism name (UCP, McPAT, DaDianNao); a design referred to as "our approach" for eleven pages has no handle for reviewers to discuss it by.
• Keep the community's units: IPC and speedup for performance, MPKI for misses/branch events, KB/mm² for cost, nJ/pJ per access for energy. Converting a miss-rate percentage where the reader expects MPKI adds friction.
• "Baseline" means the unmodified reference machine; "prior work" means the best published competitor — never blur the two in prose or legends.
• Threat-model vocabulary (for security mechanisms): attacker capability, channel, and what is explicitly out of scope, stated before the mechanism, not after.

Title and abstract conventions

The title should carry the mechanism name plus the structure touched ("Ghostline: Demand-Triggered Dead-Block Reclamation for Shared LLCs") — searchable by the structure, memorable by the name. The abstract executes the four beats in miniature, each beat with a number attached; see resources/worked-examples/01-introduction.md for a full before → after.

Anti-patterns the PC flags

• Characterization-free introductions that open with "Moore's law" boilerplate.
• Mechanisms described only in prose, with no sizeable structure — unimplementable.
• Evaluation sections that tour benchmarks instead of answering ordered questions.
• Overhead numbers missing, or buried in a sentence instead of a table row.
• IPC arithmetic averaged instead of geomean'd, or speedup over an unnamed baseline.
• "Details omitted due to space" — with no appendix allowed, that reads as "details do not exist."

Revision passes, in order

1. Argument pass — do the four beats appear, in order, with a number each?
2. Claim-audit pass — every performance sentence checked for instrument label, baseline name, and metric correctness (geomean, MPKI).
3. Skeptic pass — a co-author plays the methodology-forensics reviewer (micro-review-process) and attacks configs, sampling, and baselines; fix in text what they had to ask about.
4. Compression pass — apply micro-supplementary triage until the body sits at 11 pages without squeezing tricks.
5. Format pass — the micro-submission gate table, on paper, by eye.

Output format

[Argument beats] characterization / insight / mechanism / accounting: present-missing each
[Insight sentence] <quoted from draft, or ABSENT>
[Storage budget table] present at bit-level: yes / no
[Cost line] area-storage-power-latency all quantified: yes / partial / no
[Page budget] section over/under vs table above
[Figure audit] Fig1 motivates / Fig2 rebuildable / Fig3 geomean+baseline: pass-fail
[Calibration issues] <list of simulated-vs-silicon or baseline-vagueness sentences>