run2/efficiency.json

{
  "device": "cuda",
  "batch": 256,
  "dim": 256,
  "buffer": 4,
  "lsr1_ms": 4.534037907918295,
  "lgd_ms": 0.18603801727294922,
  "lsr1_mem_gib": 0.18018007278442383,
  "lgd_mem_gib": 0.024173259735107422,
  "speedup_pct": -2337.1566491520357,
  "mem_reduction_pct": -645.3693658151692,
  "lsr1_update_params": 252555,
  "lgd_update_params": 2886912,
  "paper_reference": {
    "lgd_ms": 166,
    "lsr1_ms": 91,
    "lgd_mem_gib": 17.81,
    "lsr1_mem_gib": 14.6,
    "speedup_pct": 45,
    "mem_reduction_pct": 18,
    "lgd_full_params": 17400000.0,
    "lsr1_full_params": 10400000.0
  }
}