Bare Metal GPUs: Dedicated machines for advanced AI workloads
Seriously powerful compute, purpose-built for your most demanding AI/ML projects.
Coming Soon: Bare Metal H200 GPUs
Powered by NVIDIA H200 Tensor Core GPUs with HBM3e, Bare Metal servers deliver exceptional performance and memory for generative AI, large language models, and high-performance computing (HPC) workloads.
How our customers use Bare Metal GPUs
Lepton AI
Lepton AI provides a cutting-edge AI inference and training platform. They focus on optimizing the deployment and utilization of GPUs in various compute environments. Efficiency, reliability and ease of use are at the core of their software design.
Supermaven
Supermaven's AI code completion tool boosts coding speed and efficiency with high-quality suggestions, in-editor AI chat, and a large context window of up to 1 million tokens. Powered by their advanced LLM, Babble, it's 2.5x larger than their previous model.
Moonvalley
High-Performance Bare Metal GPUs for AI/ML Workloads
Power your projects with the flexibility of Bare Metal GPUs
Platform Reliability
Support
Pricing
Availability
Trusted by developers for fast performance and full control
Bare Metal GPUs vs. GPU Droplets
Bare Metal GPUs and GPU Droplets are suited to different workloads. GPU Droplets offer easy scalability and fast provisioning for tasks like LLM training, model fine-tuning, or inference. Bare Metal GPUs get you maximum performance and complete control over hardware, ideal for high-throughput workloads.
Bare Metal H100 Servers
GPU Droplet H100
GPU Droplet H100x8
| GPU Memory | (v)CPUs | CPU Type | Network Cards | Machine Memory | Local Storage (Boot) | Local Storage (Scratch) |
|---|---|---|---|---|---|---|
| 640 GB | 192 | 2x Intel Xeon Platinum 8468 | Nvidia/Mellanox ConnectX-7 8x400G NIC Nvidia/Mellanox ConnectX-6 Dx 4x100GE NIC | 2 TiB | 1,024 GiB NVMe (RAID1) | 35 TiB NVMe (RAID5) |
| 80 GB | 20 | 2x Intel Xeon Platinum 8458P | Nvidia/Mellanox ConnectX-6 2x100GE NIC | 240 GiB | 720 GiB NVMe | 5 TiB NVMe |
| 640 GB | 160 | 2x Intel Xeon Platinum 8458P | Nvidia/Mellanox ConnectX-6 2x100GE NIC | 1,920 GiB | 2 TiB NVMe | 40 TiB NVMe |
Bare Metal GPU Resources
What are Bare Metal GPUs?
GPU Memory bandwidth
Monitoring GPU utilization for Deep Learning
Multi-GPU on raw PyTorch with Hugging Face's Accelerate library
How to maximize GPU utilization by finding the right batch size
The Hidden Bottleneck: How GPU Memory Hierarchy Affects Your Computing Experience
Frequently asked questions
What is a Bare Metal GPU?
A Bare Metal GPU is a physical server equipped with powerful GPUs, like the NVIDIA H100. This gives you full control over the hardware so you can install your preferred operating system. Bare Metal GPUs are isolated from other servers, providing excellent security. The server can be used for AI/ML tasks like model training, fine-tuning, and inference.
What are the advantages of Bare Metal GPUs?
- Full control: Bare Metal provides dedicated, single-tenant hardware, offering full control over the operating system, software, and configurations.
- High performance: It delivers peak GPU performance, ideal for high-scale AI/ML model training, complex computations, and real-time inference.
- Customizability: You can deeply customize the hardware and software environment, enabling tailored setups like Kubernetes clusters or other custom orchestration needs.
- Resource isolation: It offers complete isolation from other users, eliminating the risks of 'noisy neighbors' and helping to ensure security and privacy.
- Scalability: Bare Metal can be configured for single-node or multi-node setups, allowing for large-scale distributed training and high-performance workloads.
How does Bare Metal work?
Bare Metal works by providing a physical, dedicated server where you can install and run your applications. This setup is particularly useful if you need stable, high-performance infrastructure for demanding workloads like AI/ML model training or custom orchestration.
What advanced use-cases does Bare Metal support?
Bare Metal supports a wide range of use-cases, including:
- AI/ML workloads: Model training, fine-tuning, and inference, especially for large-scale data processing.
- Custom orchestrations: Technologies like Kubernetes for containerized environments or other custom setups for complex applications.
- High-performance computing: Applications that require dedicated resources, such as simulations, scientific computations, and real-time data processing.