Found 1 document(s)
Page 1 of 1
-
Training under a device-memory budget relies on tensor rematerialization: an evicted activation is either swapped to host memory or dropped and recomputed. Because GPU compute has outgrown PCIe bandwidth, cost-model plan...