Skip to content

SYSTEM Cited by 2 sources

Tahan (Meta Ethernet scale-up switch)

Tahan is Meta's first open, Ethernet-based scale-up switch — the high-bandwidth, ultra-low-latency interconnect between accelerators within a rack/pod, a domain that has traditionally required proprietary technologies (NVLink, etc.). Introduced at OCP Summit 2026, Tahan is purpose-built for Meta's MTIA 300 training accelerators and is the proof point that Ethernet can serve the scale-up domain, not just scale-out.

Hardware

  • Built on the 51.2 Tbps Broadcom Tomahawk 5 ASIC.
  • Total capacity 64 × 800G; half the ports (32×800G) provide scale-up connectivity to the accelerators via the backplane.
  • Four Tahan switches interconnect 16 accelerators in a non-blocking topology.
  • Liquid-cooled, running Meta's FBOSS NOS.
  • Targets 6.4 Tbps per accelerator in the MTIA 300 system.

The key design choice: reuse the scale-out transport

Tahan reuses the same lossless RoCE transport as Meta's scale-out network — routed IPv6, PFC, AI-tuned QoS — unifying scale-up and scale-out under one NOS (FBOSS) and one transport. Instead of a separate proprietary scale-up stack, the in-rack accelerator fabric speaks the same lossless Ethernet / RoCE as the cluster fabric. This is the central thesis: one Ethernet transport across all distance domains.

Next generation: 72 accelerators/rack on 102.4T, moving to ESUN

Meta previewed the next-generation scale-up blade: a 216×400G switch on 102.4 Tbps Broadcom Tomahawk 6C silicon, where eight switches interconnect 72 accelerators per rack — a large density jump enabled by the 102.4T generation. A future generation of this blade is planned to adopt OCP ESUN as the scale-up data plane, extending open standards to the entire scale-up stack.

Testing a copper scale-up domain as a statistics problem

A companion OCP 2026 talk ("Test Methodology for Multi-Accelerator Copper Scale-up Interconnect") frames the operational reality of the dense in-rack copper fabric Tahan serves: with thousands of connections per rack, even at 99.9% yield there are roughly 10 bad links per rack, and blind-mate cartridges leave no per-link serviceability. Meta therefore treats link health as a statistical / yield problem rather than a set of isolated debugging problems — a test methodology that screens the population instead of chasing individual faults. This is the scale-up-side analogue of the signal-screening discipline Meta applies to NVLink paths in Prometheus.

Why it matters

  • Ethernet in the last proprietary stronghold. Scale-up has been the domain of proprietary accelerator interconnects; Tahan is Meta's evidence that commodity Ethernet + lossless RoCE can do it, with a path to the open ESUN data plane.
  • One transport, three domains. By reusing scale-out's RoCE for scale-up, Meta collapses operational complexity — the same FBOSS/SAI tooling, QoS classes, and transport semantics span scale-up / scale-out / scale-across.

Caveats

OCP summit announcement. The 72-accelerator ESUN-based blade is a preview; "moving toward an ESUN-based implementation with future chips" is roadmap, not shipping. Tahan itself (Tomahawk 5, 4-switch/16-accelerator) is described as shipping today.

Seen in (wiki)

Last updated · 854 distilled / 2,448 read