micro-topic-selection
GitHub辅助判断项目是否适合投稿至MICRO会议。通过机制驻留测试、成本评估及路由表，区分ISCA、ASPLOS等竞品会议，识别硬件微架构贡献，并推荐工业界轨道。
触发场景
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill micro-topic-selection -g -y
SKILL.md
Frontmatter
{
"name": "micro-topic-selection",
"description": "Use when deciding whether a project belongs at MICRO (IEEE\/ACM International Symposium on Microarchitecture) or should route to ISCA, HPCA, ASPLOS, DAC\/ICCAD, SC, or MLSys — testing whether the mechanism truly lives inside the machine, gauging the MICRO evidence bar, and weighing the main track against the new Industry Track."
}
MICRO Topic Selection
Decide venue before writing a single evaluation section. MICRO is the flagship symposium of the microarchitecture community specifically — the 2026 edition is the 59th, which places its lineage back to the 1968 microprogramming workshops. Reviewers here are calibrated to mechanisms measured in cycles, bytes of state, and milliwatts.
The mechanism-residency test
Ask where the proposed mechanism physically lives and what evidence proves it works:
- State the mechanism in one sentence without naming any software component. If you cannot — if the sentence needs "the compiler", "the OS", or "the runtime" as a load-bearing actor — the work is cross-stack and ASPLOS deserves a look first.
- Name the structure it modifies: fetch/decode/rename stages, branch predictor tables, cache arrays and replacement metadata, prefetch engines, TLBs, memory controller queues, on-chip network routers, accelerator datapaths, or the power-management unit. A nameable structure is a strong MICRO signal.
- Name the cost line. MICRO reviewers expect area (mm² or KB of storage), power, and latency overheads next to the speedup. If the contribution has no hardware cost to account for, ask why it needs to be hardware at all.
Routing table (deadlines verified 2026-07-08; recheck live pages)
| Symptom of the project | Better home | Why |
|---|---|---|
| Mechanism spans ISA design, coherence protocols, or whole-system organization | ISCA (2027: abstracts Nov 10, papers Nov 17, 2026; Atlanta, June 2027) | Broadest architecture scope |
| Same community, but your evidence matures mid-summer | HPCA (2027: papers July 24, 2026; Salt Lake City, March 2027) | Nearest-cycle sibling |
| Compiler/OS/runtime co-designed with hardware | ASPLOS (2027: April 15 or Sept 9, 2026 deadlines) | Cross-layer is its charter |
| Circuit, EDA-flow, or physical-design contribution | DAC / ICCAD | Below the microarchitecture abstraction |
| Application-scale HPC performance study | SC | Workload, not mechanism |
| ML training/serving systems without new hardware structures | MLSys | Systems-for-ML community |
| Production silicon retrospective, deployment lessons, negative results at scale | MICRO 2026 Industry Track (inaugural) | Tailored review for industry constraints |
What MICRO itself rewards
The Test of Time record is a usable fit oracle. Recent awardees (SIGMICRO ToT pages, verified 2026-07-08): utility-based cache partitioning (MICRO 2006), 3D die-stacking microarchitecture, and the CACTI 6.0 NUCA modeling work (both honored in 2025). The pattern: a general mechanism or model that other groups can adopt, not a point-solution tuned to one workload. Accelerator papers have the same shape — DaDianNao (MICRO 2014) won by making the memory locality argument for ML accelerators, a structural insight, not a benchmark victory.
Strong MICRO shapes:
- A predictor/replacement/prefetch policy with a storage budget under a few dozen KB and a mechanism explanation for why it captures the pattern.
- A security mechanism at the speculation/cache boundary with a stated threat model.
- A modeling tool or simulation methodology the community will reuse.
- An accelerator whose novelty is a dataflow or memory-hierarchy decision, defended with ablations.
Weak MICRO shapes: pure software optimizations, ISA proposals with no implementation sketch, "we ran X on Y and report numbers" characterization with no mechanism, and circuits work with no architectural consequence.
Quick self-interrogation
Q1 Mechanism sentence written without software actors? yes / no
Q2 Modified hardware structure named? <structure>
Q3 Cost line estimable (KB, mm2, mW)? yes / no
Q4 Evidence plan reaches at least cycle-level simulation? yes / no
Q5 Would the idea survive a 2x-better software baseline? yes / no
Q6 Is the primary audience microarchitects, not systems builders? yes / no
Four or more "yes" answers → draft for MICRO. Q1 or Q6 failing → run the ASPLOS/ISCA comparison seriously before investing in the MICRO evaluation stack. Record the answers in the project README — when reviews later dispute the framing, the original fit reasoning is the fastest input to the resubmission routing decision.
Borderline cases, adjudicated
Recurring gray zones and how MICRO's own record resolves them:
- Microarchitectural security (speculation leaks, cache side channels, fault-injection defenses): squarely in scope when the defense or attack is a hardware mechanism with a threat model and a cost line. If the contribution is the vulnerability disclosure or the software mitigation, the security venues (IEEE S&P, USENIX Security, CCS) are the primary audience and MICRO the secondary.
- Processing-in-memory / near-data computing: MICRO-shaped when the paper commits to a concrete device organization and evaluates against a tuned conventional hierarchy; SC/HPDC-shaped when the story is application scaling.
- Simulation methodology and modeling tools: yes — the venue's own Test of Time record honors CACTI 6.0, and McPAT is a MICRO paper. The bar is adoption potential: the tool must change what other groups can evaluate.
- Compiler-assisted hardware (hints, ISA extensions consumed by new structures): MICRO if the hardware structure does the heavy lifting and the software change is a thin enabler; ASPLOS if the interesting decisions are split across the boundary.
- GPU/accelerator scheduling in software: not MICRO, however hardware-adjacent the vocabulary — no new structure, no MICRO.
Cycle awareness
MICRO's submission deadline sits in early April (2026 cycle: abstracts March 31,
papers April 7, both 11:59 PM EDT — not AoE), with the conference in the
October/November slot (MICRO 2026: Athens, Greece, Oct 31 – Nov 4). If you are
reading this after the April deadline, micro-workflow maps the fallback lattice:
HPCA in July, ASPLOS in September, ISCA in November, MICRO again next April. Dates
above are one cycle's snapshot — reopen microarch.org before committing a plan.
Output format
[Venue fit] MICRO main / MICRO industry track / re-route to <venue>
[Mechanism sentence] <one sentence, hardware actors only>
[Structure modified] <named pipeline/cache/predictor/accelerator structure>
[Cost line] <area, storage, power estimate or "cannot estimate — warning">
[Evidence plan] analytical / trace-driven / cycle-level sim / RTL / FPGA / silicon
[Nearest competing venue] <venue + one-line reason it loses>
[Next deadline to target] <venue, date, source URL to re-verify>
版本历史
- 9f86f09 当前 2026-07-19 16:58
