Broadcom Sian3 | BCM83628 Broadcom Sian3 | BCM83628 Item Spotlights
NVIDIA/Mellanox MMS4A00-XM (980-9IAH1-00XM00) Compatible 1.6T 2xDR4/DR8 InfiniBand XDR Silicon Photonics Optical Transceiver Module (Twin-port OSFP224, Broadcom 3nm DSP | Sian3 BCM83628, 1310nm, 500m (OS2) , Dual MPO-12/APC, DOM, SMF, IHS/Closed Finned Top)
NADDOD OSFP-1.6T-2xDR4H Optical Transceiver is an InfiniBand and Ethernet 1.6Tb/s 2x800Gb/s Twin-port OSFP224, 2xDR4/DR8 single mode, Silicon photonics-based, parallel, 8-channel transceiver using two, 4-channel MPO-12/APC optical connectors at 800Gb/s each. The parallel single mode, data center reach 8-channel (2xDR4/DR8) design uses 200G-PAM4 modulation and has a maximum fiberreach of 500-meters using 8 single mode fibers. The main application for OSFP-1.6T-2xDR4H is linking two switches together with up to 500 meters.
Leveraging the synergy between Broadcom's 3nm DSP and self-developed silicon photonics chip, the 1.6T transceiver is designed for XDR networks. Tested with the NVIDIA Quantum-X800 Q3400-RA switch, it delivers low power consumption of 25W, a BER of 5E-14, and seamless compatibility with ConnectX-8 NICs (C8180/C8280U). Supporting CX8-equipped servers such as DGX GB300/B300 and HGX B200 for high-speed interconnects, it scales from small-to-medium AI inference/training nodes to large AI factory-class clusters, setting a new performance benchmark for 1.6T XDR interconnects.
NADDOD’s XDR solution featuring the 1.6T transceiver boosts cluster energy efficiency through high bandwidth density and an optimized link architecture. In a 512-node AI cluster, this solution cuts ~ 7 racks (assuming a 20kW, 42U rack configuration) compared to NDR solution, saving 130.4kW (≈ 16–26 GPU servers’ power). Overall, it reduces data center operational costs, simplifies cabling, enhances space utilization, and enables flexible scaling—delivering efficient, sustainable connectivity for next-gen AI data centers.
Amid the global XDR chip shortage, NADDOD leverages Broadcom's stable supply of 3nm DSPs to reliably deliver 1.6T transceivers in bulk. We also strengthen chip R&D: Our self-developed silicon photonics (validated for key performance) enables easy mass production thanks to its high level of integration, providing customers with more competitive costs and reliable delivery. Meanwhile, we deploy the EML technology route, offering greater flexibility in chip selection to support product iteration and large-scale deployment for 1.6T/3.2T and future AIDC scenarios.
The 1.6T transceiver is compatible with DGX GB300, B200, HGX, and other mainstream high-density GPU servers, fully releasing the computing power of each GPU node and comprehensively boosting overall cluster performance. Ideal for AI, HPC, and large-scale inference tasks, it features high-speed, low-latency links that ensure seamless communication—activating maximum cluster throughput and resource utilization to deliver an efficient, reliable, and scalable computing infrastructure solution for customers.
