OSPA v1.0 puts Meta, Microsoft and Alibaba behind one switch-platform spec at OCP

Each GPU gets eight 100G links, one into each of eight planes ^[7]. That is 800G per GPU in aggregate ^[10]. Lose one link and the GPU keeps seven-eighths of its bandwidth, a 12.5% loss, assuming the planes are equal ^[11].

Meta lists larger non-blocking GPU clusters than single-plane topologies as the first of four FPF advantages ^[13]. It also says FPF changes the control plane of the fabric, not only how GPUs attach to it ^[14]. The post text we reviewed does not quantify the cluster-size gain or describe the control-plane change. We'd expect that change to carry most of the adoption cost. Meta says it has shared its FPF learnings with its cloud partners ^[15].

The switch numbers check out. 128 ports at 800G is 102.4T ^[12], double the 51.2T platforms IceCube follows ^[6]. IceCube comes in a Broadcom Tomahawk6C variant and an NVIDIA Spectrum-6 variant. Both run FBOSS, which uses OCP's Switch Abstraction Interface ^[5].

MetaRoCE has open-source implementations ^[2]. FBOSS is described as open source ^[5]. OSPA is a specification submitted to OCP, and it addresses the switch platform layer ^[1]. For IcePack, IceCube and FPF, the post describes designs. The text we reviewed does not say design files are being released.

In our view, the sensible order of evaluation follows that line. Read the OSPA v1.0 text and test MetaRoCE first, because Meta calls those open. Treat the rack and the fabric as architecture write-ups until a release says otherwise. The only comparative figure in the material is the 2x capacity step over 51.2T. It describes port speed, and how a given training job will behave on it still has to be measured.

Meta invites engineers, silicon vendors, system integrators and operators to join the work ^[9].