Which GPU should you actually use for embedding workloads?
Selecting the right serving engine for your embedding model can dramatically outperform hardware upgrades, yielding up to an 11x throughput increase on the same GPU.
All
Your network volume is pinned to one datacenter, which means it also picks your GPUs. If the H100s there are taken, the options are waiting for capacity or moving 300GB to another datacenter and watching a sync script the whole way.
Model weights don't need to be pinned anywhere. You write a base model, LoRA, tokenizer, or config once, and every Pod after that just reads it. The one-data-center rule comes from how network volumes are built, not from anything about the files.
Introducing Global Volumes
Today we’re launching Global Volumes in beta: elastic, region-independent storage you can mount into a Pod in any Runpod data center. Store a model once, deploy your Pod where the GPUs are available, and access the same files at /workspace-global.
How it works
When you create a Global Volume, there’s no data center to select. Attach it when deploying a Pod and Runpod mounts it into the container at /workspace-global. Your code can access the files through ordinary file paths without changing how it loads them.
That means you can store a model once, then mount the same data wherever you deploy. If the GPU you need isn’t available in one data center, you can launch in another without first creating and maintaining another copy of your model.
Use cases and trade-offs
Global Volumes is designed for data that is written infrequently and read across deployments: model weights, LoRAs, tokenizers, configuration, and similar artifacts. Capacity is elastic, so there’s nothing to size up front or resize later. Billing is usage-based; see the pricing page for current rates and billable usage.
There is a trade-off. Global Volumes is backed by object storage and is not a replacement for high-performance regional storage. Workloads that depend on full POSIX semantics, frequent writes, concurrent writes from multiple Pods, or high storage throughput should use regional network storage instead.
Getting started
The beta is available to all users and supports Pods at launch. To create one, go to Storage → New storage → Global volume. No data center selection is required.
When deploying a Pod, open Storage → Persistent storage, set the type to Storage volume, and select your Global Volume. Its files will be available inside the container at /workspace-global.
For more details on setup, see our docs or use the in-console feedback link to reach the engineering team. If your deployment process currently includes copying or downloading the same model before you can use an available GPU, Global Volumes removes that step.
