Introducing Axiom™
The memory plane
A policy engine that manages residency across every tier of memory (HBM, HBF, DRAM, CXL, local and networked flash) and composes with the stack you already run.
By keeping session state resident below HBM instead of discarding it, Axiom™ turns memory the GPU can't hold into capacity the GPU can still use.
Always in motion
Axiom™ decides what to move, and when, using two predictions that work together.
Placement is not a one-time decision. A mover runs in the background for as long as the system is live, circulating data across tiers. As information cools it falls, demoted from hot memory toward larger, cheaper tiers. When it is needed again it rises, promoted back up ahead of the request that will use it.
Everything comes to rest at the coldest tier. The archival floor is inclusive, so a copy is captured at the bottom of the plane and nothing is lost as data ages. A block can sink all the way to tape and still climb back to HBM the moment the workload calls for it.
What the dynamic side learns feeds back into the defined model, so every run sharpens the prediction for the next one.
