brycewang-stanford HPCA Writing Style
Use this to shape an HPCA paper's prose and structure. The house style rewards a paper built around one mechanism-to-behavior claim, motivated by a measured cost and evidenced on a named instrument. Reopen the current page limit before trusting the compression advice below.
The first-page arc
An HPCA first page should carry, in order: the measured problem → why existing mechanisms miss it → the mechanism → the behavior it produces → what the numbers rest on. A reviewer should finish the abstract and first column knowing what you built, what it costs, and what it changes — before any implementation detail.
- Lead with the specific pathology and a number, not "latency is a bottleneck."
- Name each prior mechanism's specific failure, not "prior work is limited."
- State the mechanism concretely enough that a reviewer could rebuild it.
- Anchor every headline number to its instrument in the same breath.
Quantify the motivation before the fix
HPCA reviewers distrust a mechanism whose problem is assumed. Characterize the cost first — measure the wasted traffic, the stall cycles, the energy — so the mechanism answers a number. A motivation section that is all citation and no measurement is a common reject signal.
Describe the mechanism so it is checkable
The mechanism is the paper. Give the structure's size, the policy's logic, and the hardware cost. If a reviewer has to guess whether your "confidence mechanism" is a standard counter or something new, the novelty is invisible. A figure of the datapath plus a precise textual description beats prose alone.
Every number names its instrument
A speedup with no instrument is unreviewable. State whether it comes from a cycle-level simulator (which one, at what fidelity, with which validated timing models) or real silicon (which machine, in which governor/turbo/SMT state), and report per-workload results with neutral cases marked — not only the mean.
Style anti-patterns
| Anti-pattern | Why it hurts at HPCA | Fix |
|---|---|---|
| "Latency is a well-known bottleneck" | No measured cost | Lead with the specific pathology and a number |
| "A confidence mechanism" | Mechanism not checkable | Give the structure, logic, and hardware cost |
| "Outperforms in most cases" | Overclaim, no instrument | Per-workload results + named simulator/machine |
| Geomean-only results | Hides where it loses | Report the distribution; mark neutral cases |
| "et al." in the bibliography | Reject trigger (all authors required) | List every author of every reference |
| Roadmap as argument | Signposting replaces substance | Replace with the mechanism-to-behavior claim |
Compression for the page limit
The 11-page body carries the characterization, the mechanism, and the headline evidence; exhaustive per-configuration tables, extra sensitivity studies, and the artifact go to the appendix and the anonymized mirror. Compress by cutting redundant background, not by trimming author lists or the fidelity contract. Leave a little slack: the single rebuttal/revision window may require small additions, and a body already jammed to the last line has no room to answer a reviewer.
Revision pass
[First-page arc] measured cost -> gap -> mechanism -> behavior -> instrument? (Y/N each)
[Motivation] quantified before the fix? (Y/N)
[Mechanism] rebuildable from the text? (Y/N)
[Numbers] each names its instrument; per-workload reported? (Y/N)
[Bibliography] every reference lists all authors? (Y/N)
[Page pressure] slack left for the window? (Y/N)
[Top three edits] <ordered>
Recheck the current CFP: the page cap, font rule, and the all-authors reference requirement are per-edition and control the compression plan.