brycewang-stanford HPCA Topic Selection
Use this before committing a project to HPCA. The question is not "is this architecture?" but "is the contribution a measured microarchitectural cost met by a buildable mechanism whose behavior is shown on a credible instrument?" If yes, HPCA fits; if the center of gravity is elsewhere, route deliberately.
The three centers of gravity
Locate where the novelty actually lives:
- The machine — a new microarchitecture, structure, or organization is the point. Strong HPCA fit when the design is buildable and its behavior is measured.
- The mechanism — a control policy, prediction scheme, or protocol that improves an existing machine. HPCA fit when a quantified cost motivates it.
- The measurement — a characterization that becomes predictive or actionable. HPCA fit when the measurement yields a tool, not just an observation.
If the novelty is a pure algorithm with no machine consequence, or a full language/OS/runtime co-design, HPCA is likely the wrong seat.
Routing across the architecture year
| Signal | Likely venue |
|---|---|
| Measured cost → buildable mechanism, hardware-grade evaluation, ready by late July | HPCA |
| Broad "new idea in architecture," flagship framing, November-ready | ISCA |
| Microarchitecture depth, pipeline/structure focus, spring-ready | MICRO |
| Cross-layer co-design spanning architecture + PL/OS/runtime | ASPLOS |
| Deployed-system evidence rather than a research prototype | HPCA Industry track |
| Pure model/algorithm with no machine consequence | a theory or systems venue |
The calendar matters: HPCA's late-July deadline is the last major architecture gate of the summer, so timing can decide routing as much as topic does. A project that just missed a spring venue often lands naturally in the HPCA winter cycle.
Main track vs Industry track
HPCA runs an Industry track alongside the main track. Route there when the contribution is evidence from a deployed or productized system — the value is the real-world measurement and the lessons, not a novel research mechanism. Route to the main track when a new mechanism or machine is the claim.
Fit test
Ask, in order:
- Is the cost measured? Can you state the pathology this paper attacks with a number, before proposing the fix?
- Is the mechanism buildable? Could a competent architect reimplement it from your description, with a stated hardware cost?
- Is the behavior shown on a credible instrument? A cycle-level simulator with a declared fidelity scope, validated timing models, or real silicon with captured state?
- Does the contribution travel? Would it generalize beyond one workload or one configuration?
Two or more "no" answers is a signal to sharpen the framing or re-route.
Worked routing example
[Project] A confidence-gated filter that suppresses useless irregular prefetches
[Center of gravity] mechanism (motivated by a measured bandwidth-waste cost)
[Cost measured?] yes — useless-traffic fraction on bandwidth-bound phases
[Buildable?] yes — one small table + gating logic, cost stated
[Instrument] cycle-level OoO sim + validated DRAM model
[Route] HPCA main track; if it were a deployed CPU's shipping prefetcher study -> Industry track
Output format
[Center of gravity] machine / mechanism / measurement
[HPCA fit] strong / plausible / weak
[Track] main / industry
[Fit-test failures] <which of the four questions answered "no">
[Recommended venue] HPCA / ISCA / MICRO / ASPLOS / other
[Reasoning] <one or two sentences>