> For the complete documentation index, see [llms.txt](https://docs.nuvolos.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.nuvolos.com/reference/applications/application-resources/fractional-gpus-on-nuvolos.md).

# Fractional GPUs on Nuvolos

## GPU Slicing on Nuvolos: Device sharing and resource isolation

Allocating a full GPU to a single user is often inefficient. In teaching and foundational research, a workload might only use a fraction of a device's actual capacity, leaving the rest of the resource idle while the institution incurs full costs.

Nuvolos GPU Slicing splits a single physical device into multiple isolated devices, going beyond solutions such as NVIDIA MiG. This allows multiple users to share the same hardware simultaneously without interfering with each other's work.

### How It Works For You

#### 1. For Students

* Dedicated Resources: Your Application runs on a dedicated, isolated computing device (GPU, NPU, or other accelerators), with a fixed allocation of the device’s capacity, i.e., cores and video  memory (VRAM).
* Workload Isolation: Your compute environment is sandboxed. If another user on the same physical card runs code that crashes or overloads their system, your workspace remains stable and unaffected.

#### 2. For Educators & TAs

* **Predictable scaling**: Deploy uniform Applications with dedicated, isolated computing devices (GPUs, NPUs, or other accelerators) to 10 or 150+ students simultaneously. Resources are sourced by Nuvolos from geo-diverse locations, ensuring simultaneous availability of the desired computing devices to all participants.
* **Consistent environments**: Every student operates on the identical hardware and software profile, eliminating environment-related troubleshooting during labs.
* **Cost constraints**: Set automated runtime limits and per-student credit quotas. If an Application is left idle over a weekend, Nuvolos automatically shuts it down to protect your budget within a time window configurable by you.

#### 3. For Research

* **Higher resource utilization**: Run up to eight concurrent workloads on a single physical device during active course hours, maximizing the return on your hardware investment.
* **Provider independent**: The Nuvolos control plane sits above the infrastructure layer. The end-user experience remains identical whether your compute resources run on Azure, SWITCH Cloud, or on-premise servers.
* **Clear telemetry**: Track usage metrics by user, department, or course to convert aggregate cloud bills into clear operational data.

### Traditional Allocations vs. Nuvolos GPU Slicing

| **Feature**          | **Traditional Allocation**                       | **Nuvolos GPU Slicing**                         |
| -------------------- | ------------------------------------------------ | ----------------------------------------------- |
| User Capacity        | 1 User per device                                | Up to 8 Users per device                        |
| Hardware Constraints | Often requires specific enterprise-grade devices | Fits across varied device tiers                 |
| Stability            | Shared environments risk full-node crashes       | Strict memory boundaries protect adjacent users |
| Cost Management      | High risk of idle runtime costs                  | Enforced auto-stop timers and credit limits     |

<figure><img src="/files/GhNrDYYsgwl59i8xXyXY" alt=""><figcaption><p>Traditional vs. Fractional device allocation</p></figcaption></figure>

> #### Behind the Scenes
>
> To deliver this, Nuvolos transforms raw bare-metal infrastructure into an expandable platform that operates uniformly across multiple cloud providers. Within this flexible cloud environment, we integrate a toolkit to dynamically slice physical devices, delivering isolated processing power straight to individual Applications regardless of where the underlying hardware physically sits.\
> \
> The system injects a custom scheduler into the Nuvolos control plane. When an Application container requests a fraction of a device, our runtime library intercepts memory allocations at the user-space boundary. If an Application attempts to exceed its assigned allocation, the library blocks the overflow, keeping the underlying node and all adjacent users completely stable.

### Join the Early Access Program

GPU Slicing is currently in beta testing to ensure stability, performance boundaries, and multi-cloud compatibility.

While this feature is not yet live in the standard Nuvolos dashboard, we are opening a waitlist for universities, research institutes, and educators who want to participate in our upcoming beta phases.

#### Why join the waitlist?

* Get early access to beta testing templates.
* Work with our engineering team to align the feature with your existing infrastructure (including major cloud providers and on-premise).
* Secure priority deployment for your upcoming course cohorts.

You can join the waitlist using the following link: [Waitlist Form](https://2f4lzm.share-eu1.hsforms.com/2ZIGzH7TCSBGvH8rKWhu6Nw)