RAM · random-access memory
Volatile working memory; in AI nodes the term usually means host RAM, which stages data-loader batches and holds the fastest checkpoint tier alongside the GPU's own HBM.
Current numbers
1.5x vs 3xstorage overhead: erasure coding (EC 4+2) vs triple replication — capacity efficiency at the cost of CPU/RAM and rebuild
~$13.6k/servercombined host-BOM trim on a training server — 2 TB → 1 TB RAM plus a host-CPU downgrade and dropping two BlueField-3 DPUs; RAM the trainer never used