
Reload or Recompute? A Calibrated Rule for KV-Cache Reuse Across Commodity GPUs
Tests when reloading a conversation’s saved KV cache is faster than recomputing it, using a break-even bandwidth rule across L4, A10 and A10G.
Frozen predictions correctly classified 574/636 resolved decisions on L4 after about a minute of calibration.

