AI Center Considerations

On top of the core service requirements that are part of full platform requirements, AI Center requires additional resources, depending on the models that you want to run or train. For more details about the required GPU hardware generations and compatible NVIDIA drivers, refer to Compatibility Matrix.

The following table describes the additional resources AI Center needs. In this table, Data Disk is needed on all server nodes. This is not needed on agent nodes.

Use	CPU	RAM (GiB)	GPU	Disk (GiB)
Minimum for serving (ML Skill, one replica)	0.6	2	0	20 GiB on the rancher partition
Minimum for Training (Pipeline)	1	4	0	20 GiB on the rancher partition Minimum 51 GiB AI Center disk partition(recommended 105 GiB)
DU model Serving (ML Skill, one replica)	1	4	0	20 GiB on the rancher partition
DU model Training	2	24	Strongly recommended	20 GiB on the rancher partition Minimum 51 GiB Data Disk (recommended 105 GiB)

The following table describes the required resources for small and average AI Center implementations. Note that these numbers are general guidance. In this table, Data Disk is needed on all server nodes. This is not needed on agent nodes.

Use	CPU	RAM (GiB)	GPU	Disk (GiB)
Small implementation: 3 models served 1 concurrent pipelines	4	32	0	80 GiB on the rancher partition 1 105 Data Disk 2
Average implementation: 5 models served 2 concurrent pipelines DU model training	8	52	Strongly recommended	160 GiB on the rancher partition3 315 Data Disk 4

¹ (3 skills +1 pipeline) * 20 GiB on the rancher partition = 80 GiB on the rancher partition

² 1 pipeline * 105GB = 105 Data disk

³ (5 skills + 2 pipeline + 1 DU pipeline) * 20 GiB on the rancher partition = 160 GiB on the rancher partition

⁴ (2 pipeline + 1 DU pipeline) * 105GB = 315 Data disk