brycewang-stanford HPCA Related Work
Use this to position an HPCA paper against decades of computer-architecture literature. The bar is a related-work section that reads as a lineage with structural deltas, not a citation pile — and one that respects HPCA's two bibliography rules: references are never blinded, and every reference lists all of its authors.
Sweep the whole architecture record
Architecture ideas recur across four venues and forty years. Search the publisher-neutral spine (dblp) plus IEEE Xplore, not a single library:
- dblp HPCA series (
dblp.org/db/conf/hpca/) for the venue's own line, and the ISCA/MICRO/ASPLOS series for cross-venue lineage. - IEEE Xplore for HPCA's published volumes; ACM DL for the ACM-published siblings.
- The HPCA Test of Time Award list for the load-bearing ancestors of your subfield.
A mechanism you think is new often has a named ancestor at HPCA 2003 or ISCA 2006; find it before a reviewer does.
Write lineage, not a pile
Each related-work paragraph should name a prior line, then state the structural delta your paper adds — what changes in the machine or the mechanism, not just that you are "different." Group by mechanism family, not chronologically.
| Weak related-work move | Stronger HPCA move |
|---|---|
| "Many prefetchers exist [3,7,9,12]." | "History prefetchers [X] cover irregular patterns but issue regardless of bandwidth; we gate on measured utilization." |
| Listing venues and years | Naming the mechanism each prior work uses and where it breaks |
| Omitting a famous ancestor | Citing it and stating the concrete delta |
| Blinding your own prior work out | Citing it in the third person, un-blinded |
Blind the voice, never the bibliography
Double-blind means the prose must not reveal you, but the references must stay intact:
- Cite your own prior work in the third person ("Prior work [12] proposed …"), never remove it.
- Do not anonymize, omit, or "et al."-abbreviate any reference — HPCA requires each reference to list all its authors, and a missing citation reads as either a gap or a blinding hack.
- Keep acknowledgments and funding out of the submitted version; they return at camera-ready.
The concurrency and preprint question
Recent HPCA guidance asks PC members not to hold arXiv or CAL preprints against a submission. Position honestly against concurrent work you know of, but do not use a preprint to claim priority in a way that de-anonymizes you. When a truly concurrent result exists, name the difference in scope rather than ignoring it.
Build sequence
1. dblp + IEEE Xplore sweep across HPCA/ISCA/MICRO/ASPLOS for the mechanism family
2. Pull the Test-of-Time ancestors of the subfield
3. Draft one paragraph per mechanism line, each ending in a structural delta
4. Blind self-citations to third person; confirm none were dropped
5. Verify every reference lists all authors (HPCA reject trigger otherwise)
Output format
[Coverage] mechanism families cited / families that exist
[Lineage quality] deltas stated / paragraphs
[Missing ancestor risk] <named prior work not yet cited, or none>
[Blinding] self-cites third-person? bibliography intact? (Y/N)
[Author-list check] every reference lists all authors? (Y/N)
[Top gaps] <ordered>
Reopen the current CFP for the preprint policy and the all-authors rule before finalizing — both are per-edition.
Source: brycewang-stanford/Awesome-Journal-Skills → HPCA-Skills/skills/hpca-related-work/SKILL.md