Redfish API

Redfish is built on a modern RESTful (Representational State Transfer) interface, using HTTPS for transport and JSON for data serialization. This makes it inherently human-readable and easily scriptable with standard tools like curl or Python requests, unlike the binary protocols of legacy IPMI. Every resource—from a chassis to a GPU—is addressable via a unique URI, enabling programmatic discovery and management of the entire hardware fleet.

Security is a foundational requirement, not an afterthought. Redfish mandates TLS 1.2 or higher for all communication, ensuring encryption in transit. Authentication is handled through a centralized Session Management model using standard HTTP Basic or Token-based auth. Role-Based Access Control (RBAC) allows administrators to define granular privileges, restricting a monitoring tool to read-only access while granting firmware update permissions only to a provisioning service.

The data model is organized around physical and logical computer systems. Key resource types include:

• ComputerSystem: The server itself, including CPU, memory, and GPU inventory.
• Chassis: The physical enclosure, power supplies, and thermal sensors.
• Manager: The Baseboard Management Controller (BMC) providing the Redfish service.
• UpdateService: A standardized endpoint for orchestrating firmware updates across multiple components. This model allows a single API to manage everything from a single GPU to an entire rack.

Redfish eliminates the need for constant polling through its Event Service. Clients can subscribe to a stream of asynchronous events pushed via Server-Sent Events (SSE) or WebSockets. This enables near-real-time alerting for critical hardware faults, such as a GPU exceeding its thermal threshold or a redundant power supply failure. The MetricReport resource provides standardized telemetry data, allowing integration with monitoring platforms like Prometheus for historical trend analysis.

The API is designed for infrastructure-as-code workflows. Using the CompositionService, administrators can dynamically assemble a server from disaggregated pools of compute, GPU accelerators, and storage. A single POST request can attach a specific GPU model to a blade and boot it with a predefined BIOS configuration. This programmatic composability is essential for bare-metal orchestration platforms like MAAS or Metal3, enabling rapid repurposing of GPU nodes for different AI training jobs.

Managed by the DMTF (Distributed Management Task Force), Redfish is an open industry standard with broad adoption across all major server OEMs. While the core schema ensures interoperability, vendors can publish their own OEM extensions to expose unique hardware features—such as NVIDIA's GPU-specific telemetry or custom liquid cooling pump status—without breaking the standard model. This provides a single, unified management interface across a heterogeneous fleet of AI infrastructure.

The Intelligent Platform Management Interface is the legacy predecessor to Redfish. It defines a set of common interfaces for out-of-band hardware management, typically operating independently of the host OS. Unlike Redfish's modern RESTful JSON approach, IPMI relies on older, less secure remote console protocols and binary message formats, making it less suitable for large-scale, automated hyperscale environments.

The BMC is the physical microcontroller on a server motherboard that exposes the Redfish API endpoint. It operates independently of the host CPU, with its own dedicated power and network connection. The BMC monitors physical sensors for temperature and voltage, controls power states, and provides the virtual KVM and media redirection capabilities essential for remote bare-metal provisioning.

Data Center Infrastructure Management software acts as the central aggregation layer that consumes Redfish telemetry from thousands of nodes. These platforms translate raw Redfish data into actionable dashboards for capacity planning, environmental monitoring, and asset lifecycle management. They leverage Redfish's event subscription mechanism to react to hardware faults in real-time without polling.

Platform Level Data Model (PLDM) and Security Protocol and Data Model (SPDM) are companion DMTF standards that work alongside Redfish. PLDM provides an efficient low-level data transfer mechanism for firmware updates and monitoring. SPDM enables hardware identity authentication and secure session establishment, ensuring the BMC you are talking to is not a compromised imposter.