As industries across the globe adopt data-intensive technologies like AI, machine learning, and real-time analytics, the demand for GPU-accelerated cloud computing is growing fast. Not every platform can deliver the performance and flexibility these workloads demand. FishOS by Sardina Systems changes that, making it easy for operators to build and manage scalable, GPU-powered clouds.
GPU-accelerated cloud computing leverages the power of Graphics Processing Units (GPUs) to dramatically speed up workloads that require large-scale parallel processing. Compared to CPUs, GPUs are optimized for handling thousands of tasks at once, making them ideal for high-performance computing tasks like training AI models, running simulations, or performing large-scale data analysis.
As the need for real-time insights and high-speed computing grows, more businesses (and not just tech giants!) are adopting GPU-powered cloud platforms to meet these demands with greater performance and flexibility. From AI/ML model training and large-scale data analytics to scientific simulations and real-time rendering, many industries now depend on GPUs to process large volumes of data quickly and reliably.
The core requirement for these businesses is simple: provide users with access to hardware in a flexible, reliable way, at an optimal cost. Whether running on bare metal, virtual machines, or containers with Kubernetes, it’s crucial to use GPU power without slowdowns.
Achieving GPU Cloud Performance Without Licensing Headaches
NVIDIA has established itself as a leader in GPU infrastructure, particularly in AI and high-performance computing. Its ecosystem, supported by frameworks like CUDA, cuDNN, and the Deep Learning SDK, is widely adopted across industries. However, e.g., CUDA-based applications can run across different deployment models, such as PCI passthrough or virtualization, without being tied to specific NVIDIA licensing structures.
While NVIDIA’s hardware is renowned for performance, its software licenses, especially for virtualization, can be costly and occasionally less productive, especially in large-scale environments where licensing issues may lead to degraded GPU performance if not managed properly. For example, if a virtual machine fails to acquire a valid license, GPU performance can degrade after 20 minutes, and CUDA functionality may be disabled after 24 hours. Although NVIDIA may improve these issues in the future, service providers must manage these risks today.
In practice, many cloud service providers, particularly those focused on reliability, seek alternatives to minimize dependency on licensed virtualization software. For instance, dedicated GPU access through PCI passthrough offers a direct, cost-effective, and robust way to assign GPU resources to virtual machines or containers. This approach doesn’t require NVIDIA’s licensed vGPU software.
FishOS cloud management platform offers compatibility with NVIDIA software, allowing users to take advantage of NVIDIA’s GPU acceleration tools if needed. However, FishOS doesn’t rely on NVIDIA software to deliver a full GPU-accelerated cloud experience.
A Hardware-Agnostic, GPU-Friendly Platform
FishOS by Sardina Systems is designed to be hardware-agnostic. Whether you are working with NVIDIA, AMD, or other vendors, this cloud solution can effectively manage GPU workloads.
While FishOS doesn’t replicate every single feature from third-party cluster managers like NVIDIA FleetCommand or Bright Cluster Manager, it provides equivalent alternatives through a powerful mix of open tools and native integration:
FishOS Capabilities and Integrated Tools
| Functionality | Technology Used |
|---|---|
| VM management | OpenStack Nova |
| Bare metal provisioning | OpenStack Ironic |
| Kubernetes deployment | OpenStack Magnum |
| Support for AI/ML | TensorFlow, PyTorch, Kubeflow, Jupyter, etc. |
| Remote nodes’ deployment | FishOS Deployer |
This modular and open approach means that even if a specific third-party feature is missing, FishOS offers flexible, reliable ways to achieve the same goals.
By integrating open source tools like OpenStack and Kubernetes, FishOS offers operators a flexible and robust platform to build and manage GPU-powered environments on bare metal, VMs, or containers:
- Nova enables GPU usage in VMs via PCI passthrough.
- Ironic provisions bare metal servers with direct GPU access.
- Magnum supports Kubernetes clusters that can run GPU workloads at scale.
This flexibility means operators don’t have to lock themselves into one vendor or ecosystem. Instead, they can build their stack based on performance, reliability, and cost efficiency.
Many popular platforms, such as Red Hat OpenShift or Canonical’s Ubuntu OpenStack, use containerized OpenStack deployments to offer flexibility and application portability. This setup suits environments where workload agility and scaling are key priorities.
However, for service providers focused on delivering dedicated GPU access, high performance, and uptime, a bare metal deployment can be a better fit. FishOS can deliver OpenStack directly on bare metal, without adding a container layer. This approach reduces system complexity, minimizes overhead, and improves reliability — all critical for operators aiming to optimize performance and lower operating costs.
When providing bare metal GPU rentals, for example, FishOS with Ironic enables direct hardware management without the need for container orchestration.
GPU Deployments That Drive Results: Full Flexibility and Control with FishOS
FishOS customers around the world leverage GPU resources via PCI passthrough, avoiding licensing constraints while maintaining performance and flexibility.
VMAccel, a global AI infrastructure provider, runs a FishOS-managed environment of 19 compute nodes with both GPU and FPGA technologies. Their cloud supports demanding workloads in AI, machine learning, genomics, and finance, where virtual machines are continuously launched and retired to match dynamic demand. FishOS ensures high uptime, scalability, and operational agility across this fast-moving environment.
Another example is the University of Tübingen, one of Germany’s leading research institutions. Their FishOS-powered infrastructure supports GPU workloads on both hypervisors and bare metal, enabling critical projects in medical imaging, cancer research, brain activity analysis, climate simulation, and more. To optimize performance, GPU nodes are prioritized for intensive tasks, while CPU nodes handle dataset management and other standard functions. Their hybrid architecture integrates virtualization and bare metal deployments, supported by Ceph storage for archiving.
Sardina Systems’ platform also supports hierarchical multi-tenancy, enabling service providers to host multiple companies, departments, and users with granular access control and resource quotas at every level.
With FishOS, service providers don’t need to compromise between performance, cost, and flexibility. Whether building a cloud for AI research, offering GPU-powered virtual desktops, or launching HPC-as-a-Service, FishOS gives you the tools to:
- Run any GPU vendor (NVIDIA, AMD, and more)
- Avoid license limitations and reduce TCO
- Deploy at scale, without added platform complexity
Talk to our experts and discover how FishOS can help you deploy a high-performance, cost-efficient GPU cloud. Book a free consultation today.