Unified Virtual Addressing
In plain English
UVA lets supported host and device allocations share an address space that GPU code can use.
Technical definition
Unified Virtual Addressing provides a unified virtual address space for supported CPU and GPU memory allocations. It does not make host memory physically equivalent to HBM.
Engineering details
In the article, the Engram kernel reads pinned host memory directly through UVA and performs row selection and dequantization on the GPU. This avoids the explicit CPU row-ID round trip used by the experimental SSD-backed implementation.
Why it matters
A common address space does not imply automatic page migration or identical access bandwidth. Performance still depends on the memory backing the address and the interconnect used to reach it.
How to read it in InferenceX
The HBM and DRAM variants use the same row-selection kernel and both support full decode graphs. The SSD experiment changes that execution path, so its comparison cannot be interpreted as a measurement of storage-device latency alone.