AI/HPC Data Center & Cloud

Where the OCS lives — the data-center fabric. Large AI/HPC clusters still rely heavily on electrical switch fabrics, where spine stages normally involve optical-to-electrical-to-optical conversion and the associated power and configuration overhead. A MEMS-based OCS from TALFOD replaces selected electrical spine stages with a transparent optical layer: micro-mirrors steer each incoming fiber straight to an outgoing fiber, creating a direct end-to-end optical path with no OEO conversion and no store-and-forward.

In a Dragonfly topology the OCS sits at the spine, forming a full mesh with N+1 redundancy so pods join or leave the fabric on demand. Because the optical path is transparent to protocol and data rate within the device’s operating bandwidth, one physical fabric can serve today's 400G/800G transceivers and later upgrades — and re-provisions in milliseconds instead of re-patching fiber in the colo.

Single-fibre bidirectional links — fewer fibers and ports. An OCS link should use as little fiber as possible. Our planar and arrayed circulators let a single fiber carry both directions: transmit and receive are routed through different ports inside the device. High isolation helps limit crosstalk, while low insertion loss preserves the link budget and reduces the number of fibers and OCS ports required.

From discrete 3-port devices to arrayed integration, we cover the full circulator range an OCS link needs. The 8-in-1 arrayed circulator packs eight bidirectional paths into one high-density package for dense OCS ports. At the far end, a 2D optical fiber collimator array couples a high-port OCS into a compact grid, while fiber shuffle chassis re-map high-count links between the OCS and its line cards.

Multi-plane fabrics and the Shuffle Box. Beyond OCS, TALFOD also supplies Shuffle Boxes and flexible boards for multi-plane AI networks. Multi-plane networking assigns the different optical channels of the same server to separate, independent network planes, so a cluster no longer scales only by adding switching tiers. At the physical layer, the Shuffle Box organizes and re-maps fibers according to a predefined mapping, so servers connect to several planes in order, larger GPU counts can be attached, and training and inference get multiple transmission paths. A pre-designed fiber mapping also reduces manual patching and cabling errors and simplifies later expansion and maintenance. Traffic scheduling and failover are handled by the network interface cards and the switching fabric; when part of a path fails, the remaining paths keep carrying traffic, improving the reliability of training and inference networks.

MCF FIFO — denser optical interconnects for AI clusters. As AI data centers accommodate growing numbers of optical channels, space inside equipment and cabling management become increasingly challenging. Our MCF FIFO (fan-in / fan-out) connects four single-mode fibers to the four cores of one multi-core fiber, using space-division multiplexing to carry four optical channels over a single fiber. Its miniature, low-crosstalk design saves cabling space in high-density optical modules and inside equipment, while simplifying connections and cable management for AI cluster expansion.

OCS as test equipment — one switch, shared by every instrument. In a lab or on a production line, devices under test (DUT) and the instruments that measure them — light source, BERT, power meter, any instrument — are normally re-cabled by hand every time the test plan changes. Put a non-blocking OCS in between, and any instrument can be software-routed to any DUT on the opposite side: one OCS shared by every test station.

Instrument ports and DUT ports line up one to one, and the crossing paths are built and torn down in software, so the test topology is configured once and re-used across many devices. It is the same optical-switching core that drives a data-center fabric — here it switches test resources instead. Our OCS is developed fully in-house, including the core 2D FAU collimator arrays, so the any-to-any paths hold stable insertion loss and stay reliable as long-term test equipment.