Optical interconnect for AI infrastructure
AI / GPU cluster fabric
Spine-leaf and rail-optimized fabrics for GPU training clusters using 400G and 800G DR/SR optics with breakout.
Data center interconnect
FR4 / LR4 optics for campus and metro DCI links up to 10 km, connecting availability zones and pods.
Low-power & LPO
Silicon-photonics and linear-drive (LPO) options to cut power and latency in dense high-speed fabrics.
Spine-leaf fabric for AI clusters
Twin-port OSFP fabrics, end to end
Two straight single-mode fibers, one twin-port transceiver, two endpoints — the wiring model behind current 400G and next-generation 800G InfiniBand clusters. PhoScale supplies the optics and cabling for both generations.
Twin-port 2×400G OSFP fabric
A twin-port 2×400G OSFP switch drives two independent 400G endpoints from a single cage. Our twin-port OSFP 2×400G single-mode transceiver (OSFP finned-top, dual MPO-12/APC) fans out over two straight single-mode fibers — each carrying 4×100G-PAM4 — to ConnectX-7 (400G OSFP) and BlueField-3 (400G QSFP112) endpoints. No breakout cassettes, no polarity gymnastics: straight MPO-12/APC patch, 1–500 m.
Quantum-X800 · twin-port 2×800G OSFP224 fabric
The same twin-port model, doubled. On an NVIDIA Quantum-X800 InfiniBand switch, our twin-port OSFP 2×800G single-mode transceiver (OSFP224 closed-top, dual MPO-12/APC) links two 800G endpoints over two straight single-mode fibers — each now carrying 4×200G-PAM4 — to ConnectX-8 OSFP224 hosts. Doubling the lane rate keeps the fiber plant and connector footprint identical to the 400G generation, so the cabling you deploy today carries forward.
Same fiber, same connector, twice the lane rate — a straight-fiber MPO-12/APC plant built for 400G InfiniBand carries the 800G generation without re-cabling. Request an RFQ to match transceivers and fibers to your switch and NIC platforms.