Skip to content

M6_NOTES.md

Latest commit

 

History

History
26 lines (21 loc) · 1.06 KB

M6_NOTES.md

File metadata and controls

26 lines (21 loc) · 1.06 KB
Raw

M6 — Skip extreme cases by default

Decision

Default behavior now skips (C,z,x) cases where A_max > 10M, and y-slots where B_max > 10M. This cuts ~95% of GPU work from the M5 profiling histogram tail.

Trade-off

Loses hits with one of A,B > 10M. For bound=500 e3-10 raw, this drops some hits (e.g. 196690860^3 + 68841801^3 = 309^10) from default coverage.

Flags

  • (default) : skip A_max,B_max > 10M
  • --max-Amax=N : skip A_max,B_max > N
  • --max-Amax=100000000 : skip > 100M (medium coverage)
  • --exhaustive : disable skip entirely (full coverage, slow)

Performance estimate

  • bound=500 e3-10 raw (default 10M): ~5-10s (vs 100s exhaustive)
  • bound=500 e3-10 raw --exhaustive: ~100-120s (full M3.9 baseline)
  • bound=1000 e3-12 raw (default): ~30-90s
  • bound=1000 e3-12 raw --exhaustive: probably 2-3 hours

Workflow

  • Calibration / dev: default 10M (fast, almost-complete coverage)
  • Production overnight runs: --exhaustive for complete coverage
  • Mid-range exploratory: --max-Amax=N for tunable trade-off