Pinned host memory
Also known as page-locked memory
In plain English
Pinned host memory stays resident in RAM so the device can use a stable host-memory allocation.
Technical definition
Pinned host memory is host RAM held resident for device access or transfer, rather than ordinary pageable memory that the operating system may reclaim or move through paging.
Engineering details
The Engram DRAM implementation reads a pinned host table directly through UVA. The SSD experiment instead gathers selected rows into pinned buffers before copying them to the GPU. Both use pinned memory, but they expose different coordination and transfer paths.
Why it matters
Pinned memory consumes host capacity and should not be confused with free storage. Its stable residency can support efficient device access, while excessive allocation reduces RAM available to applications and the filesystem cache.
How to read it in InferenceX
A warm SSD-backed file is not equivalent to the pinned DRAM table in the article. Cached pages avoid physical disk reads, but the experimental path still gathers rows on the CPU and transfers the resulting buffers between graph segments.