Documentation Index
Fetch the complete documentation index at: https://docs.poolside.ai/llms.txt
Use this file to discover all available pages before exploring further.
Overview
Poolside officially tests and supports a defined set of deployment configurations for enterprise environments. Use this page to compare supported deployment paths, review the minimum platform, compute, database, and inference requirements, and find the configuration that best matches your infrastructure.Platform requirements
For architecture, installation, and operational guidance for supported cloud deployment paths, see Cloud deployment. For on-premises configurations, see On-premises deployment.If you already run the Terraform-based AWS deployment bundle, see Amazon EKS with Terraform (legacy) and the migration guide.
| Deployment | Description |
|---|---|
| Amazon EKS | Amazon Elastic Kubernetes Service (EKS) environments |
| OpenShift 4.16+ | Red Hat OpenShift environments |
| Kubernetes 1.29+ | Self-managed Kubernetes environments such as RKE2 or Charmed Kubernetes |
Compute requirements
The following minimum requirements apply to the Poolside Platform plane. These figures exclude model inference and remote execution sandboxes, which you size separately based on your workload.| Resource | Minimum |
|---|---|
| CPU | 24 cores |
| Memory | 64 GB |
| Storage | 200 GB |
Database requirements
You provision and manage the PostgreSQL database. Poolside does not provision it on your behalf. The following sizing is a recommended starting point:| Resource | Recommended |
|---|---|
| CPU | 4 cores |
| Memory | 16 GB |
| Storage | 100 GB |
Inference requirements
Poolside models have different minimum requirements. Use this table to size your inference nodes. For concurrent-agent capacity and developer seat estimates by hardware tier, see Capacity planning. Laguna is the recommended primary model family for new deployments.| Model | Quantization | Minimum GPU memory | Minimum CPU | Minimum host memory |
|---|---|---|---|---|
| Laguna M.1 | FP8 | 384 GB | 128 cores | 1 TB |
| Laguna XS.2 | FP8 | 96 GB | 44 cores | 512 GB |
| Malibu 2.2 | FP8 | 192 GB | 128 cores | 1 TB |
| Malibu 2.2 | INT4 | 96 GB | 44 cores | 512 GB |
| Point | FP8 | 96 GB | 44 cores | 512 GB |
Certified GPUs
All models are continuously certified in the following NVIDIA GPU types: RTX 6000, H100, H200, B200, B300.Standard on-premises configurations
For the certified single-node on-premises software stack and version inventory, see Certified stacks.| Deployment | Recommended scale | GPU configuration | Host OS |
|---|---|---|---|
| BYO hardware | Large enterprise teams | 8× NVIDIA H200 (HGX) (4× H200 with validation) | Ubuntu 22.04 LTS, Ubuntu 24.04 LTS, or RHEL 9.6 |
| Turnkey HGX rack | Large enterprise teams | 8× NVIDIA H200 (HGX) | Ubuntu 22.04 LTS, Ubuntu 24.04 LTS, or RHEL 9.6 |
| GPU workstation tower | Small teams and individual groups | 4× NVIDIA RTX 6000 | Ubuntu 22.04 LTS, Ubuntu 24.04 LTS, or RHEL 9.6 |
| GPU workstation rack | Mid-sized teams | 8× NVIDIA RTX 6000 | Ubuntu 22.04 LTS, Ubuntu 24.04 LTS, or RHEL 9.6 |
For on-premises hardware deployments:
- Inference and platform services run on the same node. The node must meet both the inference requirements above and the platform compute requirements.
- Multi-node inference is supported across a single Kubernetes cluster, but it does not provide high availability. It distributes independent inference replicas across nodes but does not support cross-node tensor parallelism.
If you deploy on Red Hat Enterprise Linux (RHEL), pin the host release to RHEL 9.6 before you run package updates. RHEL can upgrade the host to a newer minor release when new updates become available. Prevent that upgrade because NVIDIA GPU Operator driver containers are pinned to specific RHEL releases. For the required commands, see Install Poolside on-premises.