Bare Metal GPU Features

DigitalOcean Bare Metal GPUs are dedicated, single-tenant servers with 8 NVIDIA H100 GPUs that can operate standalone or in multi-node clusters.

---

Bare Metal GPUs vs GPU Droplets

DigitalOcean Bare Metal GPUs and GPU Droplets both provide compute resources tailored to AI/ML workloads, but they’re each suited for different use cases.

GPU Droplets	Bare metal GPUs
Virtual machines. GPU Droplets have the convenience and ease of deployment that comes with managed infrastructure, but VM configuration is constrained by the hypervisor and shared OS layer.	Physical servers. Bare metal GPUs are physical servers without virtualization, so you can set up advanced orchestration layers, containers, operating systems, and other deep configuration directly on the hardware.
Shared infrastructure. GPU Droplets share physical resources, so there may be minor resource fluctuations that don’t significantly impact tasks like fine-tuning and inferencing.	Single tenant hardware. Bare metal GPUs are in isolated environments, which are best for use cases requiring full data isolation or highly consistent performance.
On-demand instances with per-hour billing. Pricing for GPU Droplets is flexible and low commitment, so GPU Droplets are best for variable usage or rapid scalability.	Contract-based billing and provisioning. Pricing for bare metal GPUs is more cost effective, but meant for long-term use with intensive, prolonged workloads that need stable performance.

GPU Droplets are best for small- to medium-scale tasks, including:

• Fine-tuning (adjusting models with specific data sets)
• Inference (running predictions with high-speed responses for production applications)
• Moderate data processing (lightweight analytics or video processing that benefit from GPU acceleration but don’t demand full hardware dedication)

Bare metal GPUs are best for advanced and custom workloads, including:

• Model training at scale (training foundational models and handling large datasets with optimal performance)
• Complex inference needs (running real-time inference for high-throughput applications)
• Custom orchestration and HPC (like Kubernetes clusters, multi-node setups, or high-frequency trading)

Hardware

Bare metal GPUs have the following hardware specifications:

Name	Description	Quantity
CPU	Intel(R) Xeon(R) Platinum 8468	2
Memory	64GB	32
NVMe SSD storage	7TB 2.5-inch NVMe SSD drives	8
GPU module	NVIDIA HGX H100 80GB 700W SXM5 GPUs, fully interconnected with NVIDIA NVLink technology	8
Network card	Mellanox Network Adapter; Mellanox Technologies MT2910 Family [ConnectX-7]; link speed 400 Gbps	8
Network card	Mellanox Technologies MT2892 Family [ConnectX-6 Dx] ; link speed 100Gps	4
RoCE2 RMDA network	3.2Tbps
Public network bandwidth (North/South)	N/S internet access up to 40Gbps
Private network bandwidth (East/West)	E/W private network up to 400 Gbps (GPU interconnectivity)

Software

Bare metal GPUs come preconfigured with one of the following software options:

Software	Option 1	Option 2
Operating system	Ubuntu 22.04	Ubuntu 24.04
GPU driver	535.161.08	535.161.08
CUDA driver	12.2	12.4
SSH remote access	Preloaded SSH keys	Preloaded SSH keys