Hardware Infrastructure Requirements#
The standard deployment is a five-node cluster with two or more worker nodes.
Standard Deployment#
The compact deployment is a three-node cluster.
Compact Deployment#
Holoscan for Media also supports single-node OpenShift (SNO), although SNO lacks the control-plane redundancy that NVIDIA recommends for production.
Infrastructure Examples#
The following tables show two example bills of materials for standard and compact Holoscan for Media clusters that use the RTX PRO 6000 Blackwell Server Edition GPU. Depending on the application workloads that you run, you can vary the network fabric, the number of worker nodes, and the resources (CPU, RAM, GPU) per worker node. These examples are for worker nodes based on the PCIe-optimized reference configurations for NVIDIA-Certified Systems and the related Enterprise Reference Architecture. The control-plane node specifications follow the Red Hat OpenShift recommendations for control plane node sizing.
The following table lists the standard deployment, a five-node cluster that uses 2-8-5 systems:
# |
Part |
Type |
Quantity |
Recommendation |
|---|---|---|---|---|
1 |
Jump Node |
Physical server or virtual machine |
1 |
2 cores, 4 GB RAM, 120 GB boot drive |
2 |
Control Node |
Physical server |
3 |
8 cores, 32 GB RAM, 500 GB NVMe boot drive |
3 |
Worker Node |
Physical server |
2 |
2x 64 cores, 1024 GB RAM, 1x 1 TB NVMe boot drive, 2x 4 TB NVMe storage (adjust for your use case) |
4 |
Ethernet Switch |
Multicast and PTP-capable 200G switch |
1 |
|
5 |
NIC |
ConnectX-7 2x200G |
4 |
Per Worker Node |
6 |
DPU |
BlueField-3 2x200G (B3220) |
1 |
Per Worker Node (optional, depending on your use case) |
7 |
GPU |
RTX PRO 6000 Blackwell Server Edition |
8 |
Per Worker Node |
8 |
Cables |
8 |
Per Worker Node |
The following table lists the compact deployment, a three-node cluster that uses 2-4-3 systems:
# |
Part |
Type |
Quantity |
Recommendation |
|---|---|---|---|---|
1 |
Jump Node |
Physical server or virtual machine |
1 |
2 cores, 4 GB RAM, 120 GB boot drive |
2 |
Control-Worker Node |
Physical server |
3 |
2x 64 cores, 512 GB RAM, 1x 1 TB NVMe boot drive, 2x 4 TB NVMe storage (adjust for your use case) |
3 |
Ethernet Switch |
Multicast and PTP-capable 200G switch |
1 |
|
4 |
NIC |
ConnectX-7 2x200G |
2 |
Per Node |
5 |
DPU |
BlueField-3 2x200G (B3220) |
1 |
Per Node (optional, depending on your use case) |
6 |
GPU |
RTX PRO 6000 Blackwell Server Edition |
4 |
Per Node |
7 |
Cables |
4 |
Per Node |
Network Fabric Configuration#
You can request an example switch configuration for NVIDIA Onyx switches, such as the NVIDIA Spectrum SN3700.
The production setup manual installation assumes the following configuration:
VLAN |
Subnet |
Purpose |
|---|---|---|
100 |
10.21.10.0/24 |
Management |
200 |
192.168.20.0/24 |
Media A (red) |
201 |
192.168.120.0/24 |
Media B (blue) |
300 |
172.20.0.0/24 |
PTP |
The PTP domain is 127.