On-Demand GPU Embeddings - Testing SIE Model Switching and VRAM Eviction

Download printable cheat-sheet (CC-BY 4.0)

30 Jul 2026, 00:00 Z

60-second takeaway
SIE successfully switched between a small embedding model and a reranker on one RTX 3090 Ti. A 15-second idle policy unloaded each model, and stopping the server returned GPU memory to the 47 MiB host baseline.
This proves the tested on-demand lifecycle. It does not prove SIE's pressure-driven LRU behavior, production throughput, or large-model safety.

The practical problem

A remote embedding service can save memory on a laptop, but moving the model to a GPU workstation creates another problem: does the model occupy GPU memory all day, even when nobody is searching?

An on-demand server should do four things well:

  1. load the requested model when work arrives;
  2. tell the client when that model is still loading;
  3. switch between model types without restarting the service; and
  4. release most model memory after the model becomes idle.

We tested those behaviors in SIE with one embedding model and one reranking model. The goal was not to reproduce a broad benchmark. It was to find out whether SIE could support a practical remote retrieval service on a shared 24 GB GPU.

Test setup

The server ran on an RTX 3090 Ti and listened only on 127.0.0.1. Telemetry was disabled, and the idle eviction threshold was set to 15 seconds.

The two models were:

  • sentence-transformers/all-MiniLM-L6-v2 for embeddings;
  • cross-encoder/ms-marco-MiniLM-L-6-v2 for reranking.

The repository, Python environment, package cache, and model cache lived on an external hard drive. That protected the workstation's nearly full NVMe, but it also made cold installation and model loading slower.

What happened on CUDA

StageObserved GPU memoryWhat it establishes
Host before server47 MiBIdle host baseline
Server before a model390 MiB

Turn AI video into a repeatable engine

Build an AI-assisted video pipeline with hook-first scripts, brand-safe edits, and multi-platform delivery.