This blog is co-authored by Samer El-Haj-Mahmoud, Arm & Tim Lewis, Insyde Software.
Modern datacenters run at massive scale. Even small inefficiencies in server management can multiply into higher costs, downtime, or missed performance targets. To meet strict Service Level Objectives (SLOs), operators need continuous, reliable visibility into platform health without disrupting workloads.
This is where out-of-band (OOB) telemetry comes in. By streaming real-time insights into thermal, power, and hardware subsystems—independently of the OS—OOB telemetry enables proactive management and automation at fleet scale.
In a previous Arm community blog, we presented Arm’s work for advancing server manageability. In this blog post, we show how Arm, in collaboration with Insyde Software is advancing OOB telemetry on Arm Neoverse–based servers. Built on open standards and supported by a working proof-of-concept, this architecture is designed to scale from silicon sensors all the way to datacenter analytics. This approach ensures interoperability and readiness for production environments.
Figure 1: End-to-end OOB telemetry architecture
The above figure shows the end-to-end flow of the OOB Telemetry architecture on Arm servers. Key components are:
Figure 2: Telemetry Data Capture Format (TDCF)
As explained in a previous Arm Community blog post, SMCF normalizes on-‑chip monitors (thermal, voltage, activity, fabric, and third-party IP), schedules sampling, and writes compact samples to memory—so the SCP can serve telemetry without ad-‑hoc‑ register scraping.
On top of that, SCMI is the single contract the SCP exposes. With System Telemetry added to SCMI v4.0, agents can: discover available Telemetry Data Events, configure cadence/filters, and collect data via snapshot or buffered methods from various on-chip sources including SMCF. SCMI also defines Telemetry Data Capture Format (TDCF), as shown in Figure 2. This same data is surfaced in-band (IB) to the OS using the ACPI PCC mailbox, and out-of-band (OOB) to the MCP—identical payloads, two paths. That symmetry keeps the decoder portable and lets the platform owners decide which data are IB+OOB, versus IB only.
Figure 3: Arm SBMR architecture
As outlined in a previous Arm Community blog and in Figure 3, Arm SBMR server management architecture relies on DMTF MCTP and PLDM standards for communication over the MCP (which is a Satellite Management Controller or SatMC) and the BMC.
Figure 4: Redfish and OTel handoff
The above figure illustrates an example of a typical out-of-band telemetry deployment in datacenters. The setup uses some platform components and off-the-shelf software packages:
Figure 5: Example Redfish TelemetryData resource and the attached SCMI telemetry
Figure 6: Example SCMI TDCF telemetry data
The BMC forwards PLDM File I/O payloads to an OpenTelemetry Collector with a custom receiver that parses the SCMI TDCF telemetry container (header, records, timestamps). From there, you can export via OTLP to a database (such as Victoria Metrics in our prototype) and visualize in Grafana. For example, Socket and Core thermal envelopes, or workload-correlated power traces. The OpenTelemetry documentation shows how to build a receiver and assemble a custom Collector. Grafana documents ingesting OTLP natively.
Figure 7: Visualizing bulk telemetry data using Grafana
Figure 8 – OOB Telemetry prototype
The above figure shows a complete setup of this proof-of-concept using Arm Neoverse FVP, OpenBMC, OpenTelemetry, Victoria Metrics, and Grafana. All the source code of this proof-of-concept prototype is published here.
Insyde took the Arm prototype and made production-level modifications. First, the BMC model was updated from the Arm Base FVP to the QEMU emulation of the ASPEED AST2700. This model supports a wide variety of the standard AST2700 peripherals, and that allows usage of the production OpenBMC code from Insyde’s Supervyse OPF product. Second, for the Neoverse FVP model from Arm, Insyde transitioned to use its fully featured InsydeH2O UEFI firmware product, allowing the simulation of basic functionality and numerous extended features. Combined with the firmware images for Arm’s SCP and MCP, this provides simulation of a powerful, fully featured Arm data center system.
Figure 9 – OOB Telemetry with InsydeH2O UEFI firmware and Supervyse OPF OpenBMC firmware
To learn more about this architecture and see a demo of the Arm and Insyde software prototypes, visit the Arm booth at the 2025 OCP Global Summit 2025, San Jose, CA, Oct 13-16, 2025.