I let a server treat added DRAM as a slower tier. My memory‑expansion use case ships today: Azure’s M‑series preview on Astera Labs controllers in Feb 2026, per disclosure. Tiering is already in production at Meta, a paper shows performance within 1 % of an ideal baseline. Pooling – the ‘break the memory wall’ story – remains aspirational; the flagship study was an emulation, the 100‑TB pools are proofs‑of‑concept, and no GA multi‑rack deployment exists, so research treats it as a 2027‑plus option.
← stream@CXL needs