The Open Compute Project Foundation (OCP) has launched a network specification aimed at standardizing switch platforms.
Part of the OCP Networking Project, the Open Switch Platform Architecture (OSPA) establishes standardized interfaces and design guidelines for “the final frontier of proprietary fragmentation in open networking” – the switch platform layer.
Meta and Microsoft are among those backing OSPA, which centralizes platform management services on the host subsystem, limiting the baseboard management controller (BMC) to key supervisory functions such as power management, host bring-up, leak protection, and troubleshooting and recovery. Input/output bridge field-programmable gate arrays (FPGAs) consolidate various hardware controllers into a single, fast-access interface, simplifying communication between the host and peripheral devices.
In addition, the atomic platform entity (APE) model provides deterministic hardware inventory management through a hierarchical structure with identification programmable read-only memory requirements. Firmware reliability is enhanced through serial peripheral interface-based upgrades compatible with standard Linux tools, A/B bank redundancy and dual-path recovery, while standardized error reporting and hardware fault logging aim to simplify diagnostics and troubleshooting.
The architecture also supports autonomous power management and thermal control.
Tao Ren, software engineer at Facebook, said OSPA complements other open standards such as the switch abstraction interface (SAI), as adopted by the Facebook Open Switching System (FBOSS) and Software for Open Networking in the Cloud (SONiC), with a vendor-neutral reference architecture.
With it, the OCP aims to solve the pain point of re-implementing critical platform functions for every new hardware SKU, which creates a “profound inefficiency” where identical ASICs from different manufacturers present entirely different platforms to the operating system.
“(OSPA) enables true software reuse, simplifying and accelerating the platform onboarding process – a crucial capability in the fast-paced AI industry where rapid deployment is essential,” Ren said.
“By codifying these common platform design guidelines into a shared reference design, new switch platforms can onboard more efficiently – leveraging reusable hardware designs and readily available software stacks rather than building from the ground up.”
Alibaba worked on the standard with Meta and Microsoft, with contributions from Arista, Cisco, Wistron and others. A draft version of the spec is available via GitHub.
Several OSPA backers also previously helped develop an OCP specification for Ethernet-based scale-up networking. The aptly named Ethernet for Scale-Up Networking (ESUN) 1.0 specification aimed to tackle scale-up GPU connectivity and other AI-heavy deployments.
Comments