Conjunction Search - GPU based Whole-Satcat Closest Approach Search

Leaderboard - Fastest 24h Searches - benchmark your GPU/CPU by running a full scan

Architecture

For each bucket window [T, T+bucket]:
  1. CPU: SGP4 propagate all sats at T and T+bucket (reuse previous T+bucket as new T)
  2. GPU: SLERP interpolate positions at 1-second intervals → cell entries
  3. GPU: Radix sort entries by (time_offset << 26) | cell_id
  4. GPU: Collision detection on sorted entries → candidate pairs
  5. CPU: Refine TCA for unique pairs using linear algebra root find/rejection

Data flow (GPU buffers, no CPU roundtrip until final pairs):
  positions_t0[28k] → Interp shader → entries[28k×bucket] → RadixSort → Detection → pairs[]
  positions_t1[28k] ↗