GPUs on Hikube
Hikube offers NVIDIA accelerators attached in passthrough mode, for two types of workloads: virtual machines and Kubernetes cluster nodes. There is no GPU page in the console: you choose the GPU in the wizard of the resource that uses it.
Usage modes
GPU on a virtual machine
The physical GPU is assigned to the VM through PCI passthrough: the VM has exclusive access to it and native performance.
- Chosen in the Create an Instance wizard, Configuration step, Hardware Acceleration (GPU) section.
- One or more GPUs per VM, of one or more models.
- The NVIDIA drivers are installed in the VM's OS (see Install CUDA).
Use cases: CUDA development environments, applications requiring full control of the GPU, graphics rendering, specialized workloads.
GPU on Kubernetes
GPUs are attached to the nodes of a cluster node group, then assigned to pods through resources.limits.
- Chosen in the Create cluster wizard (or Edit), Nodes step, GPU section of the node group.
- The GPU Operator addon is enabled automatically as soon as a group has GPUs; it installs the drivers and the device plugin.
- The HAMi addon lets several pods share the same GPU.
Use cases: containerized AI/ML, inference at scale, parallel jobs.
Available models
| Model (console label) | Memory | Architecture | Typical use |
|---|---|---|---|
| NVIDIA L40S | 48 GB | Ada Lovelace | Inference, generative AI, real-time rendering, prototyping |
| NVIDIA A100 80GB | 80 GB | Ampere | ML training, fine-tuning, scientific computing |
| NVIDIA H100 80GB | 80 GB | Hopper | Training and inference of large models |
| NVIDIA RTX 6000 Pro | 96 GB | Blackwell | LLM, intensive computing |
Each card in the selector shows the model name and its memory (for example 48 GB VRAM). A model with no free unit left is greyed out and marked Unavailable.
The console only shows whether a model is available or not, without displaying the number of free units. GPUs are resources shared between the platform's customers: a released GPU (stopped VM, deleted node) can be assigned to another workload. For a specific GPU capacity requirement, contact sales@hidora.io.
Architecture
GPU on VM
A VM runs on a single physical node: all the GPUs requested for a VM must be available on the same node.
GPU on Kubernetes
Comparison
| Aspect | GPU on VM | GPU on Kubernetes |
|---|---|---|
| Isolation | GPU dedicated to the VM | GPU assigned to pods by the scheduler |
| Performance | Native (passthrough) | Native (device plugin) |
| Drivers | To be installed in the OS | Installed by the GPU Operator |
| Scaling | Vertical (edit the VM) | Horizontal (number of nodes in the group) |
| GPU sharing | No | Yes, with the HAMi addon |
| Modification | Add, remove or change GPUs (restart) | An existing group keeps at least one GPU; a group without GPUs cannot receive any |
Billing and quotas
The cost estimate shown in the VM and Kubernetes wizards includes the selected GPUs. A VM's GPU is released when the VM is stopped.