Microsoft and UC Merced present ZeRO-Offload
Presented ZeRO-Offload — offloads optimizer states to the CPU, reaching 40 TFlops for a 10B model on a single V100
Why it matters
The start of the lineage that lowered the hardware barrier to entry for training large models
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/2101.06840Added in Week 35, 2026 · See that week