NVIDIA MMS4A20-XM800 (980-9IAT0-00XM00/980-SIAY0-00XM00) Compatible 800G DR4 OSFP224 4x200G-PAM4 1310nm 500m InfiniBand XDR Module Featuring Broadcom 3nm DSP (Sian3 | BCM83628) for SMF (Riding Heat Sink/Flat Top)

Name: NVIDIA/Mellanox MMS4A20-XM800 (980-9IAT0-00XM00/980-SIAY0-00XM00) Compatible 800G DR4 OSFP224 RHS/Flat Top 4x200G-PAM4 Broadcom 3nm DSP (Sian3 | BCM83628) 1310nm SMF 500m DOM MTP/MPO-12 APC InfiniBand XDR Transceiver Module for ConnectX-9/ConnectX-8/B300/GB300
Brand: Naddod
SKU: 102622
Price: 1399.00 USD
Availability: OnlineOnly
Rating: 5.0 (18 reviews)

Description

NVIDIA/Mellanox MMS4A20-XM800 (980-9IAT0-00XM00/980-SIAY0-00XM00) Compatible 800G DR4 InfiniBand XDR Optical Transceiver Module (OSFP224, Broadcom 3nm DSP | Sian3 BCM83628, 1310nm, 500m , Single MPO-12/APC, DOM, SMF, RHS/Flat Top)

NADDOD OSFP-800G-DR4H is an 800Gb/s single-mode optical transceiver designed for InfiniBand XDR networks. It is compatible with NVIDIA/Mellanox MMS4A20-XM800 (980-9IAT0-00XM00/980-SIAY0-00XM00). The 800G DR4 transceiver operates at a line rate of 200Gb/s per lane using 4 × 200Gb/s PAM4 modulation, delivering high-speed, low-latency connectivity. For physical connectivity, OSFP-800G-DR4H supports link distances of up to 500 meters over single-mode fiber (SMF) using a DR4 optical architecture, delivering an aggregate bandwidth of 800Gb/s.

With an integrated advanced 3nm DSP, the 800G OSFP DR4 transceiver delivers ultra-low bit error rates (BER), robust signal integrity, and high end-to-end link reliability at 800G speeds. Silicon photonics technology further enhances integration density and energy efficiency. This transceiver is optimized for liquid-cooled systems based on the NVIDIA GB200 architecture, linking Quantum-X800 QM3x00 switches with ConnectX-8 dual 800Gb/s mezzanine cards via high-density 800G optical interconnects.

Specifications

Form Factor

OSFP224

Distance

500m

Wavelength

1310nm

Connector

MTP/MPO-12 APC

Power Consumption

≤15W

Transmitter

DFB

Receiver

Media

SMF

Modulation (electrical)

4x212.5Gb/s PAM4

Modulation (Optical)

4x212.5Gb/s PAM4

Case Temperature (℃)

0~70°C (32 to 158°F)

Protocol

IB XDR

Applications

Product Highlights

3nm DSP × In-house SiPh for High-Performance XDR Interconnects

In high-density XDR networks, 800G optical links between servers and switches are required to operate at full load over long periods, which places higher demands on power consumption, heat dissipation, and link stability. NADDOD 800G DR4 transceiver features Broadcom 3nm DSP (BCM83628) and in-house silicon photonics chips. This design effectively reduces power consumption and thermal burden, and significantly enhances signal quality and link reliability, building stable and dependable high-speed interconnects for large-scale GPU clusters.

Real Cluster Validation • Long-Term Stable Operation

NADDOD 800G transceivers have been delivered at scale and deployed in NVIDIA HGX B300 clusters, demonstrating stable, long-term operation. These transceivers are compatible with mainstream high-density GPU servers such as NVIDIA DGX B300/GB300 and OEM HGX B200. They work seamlessly with switch-side 1.6T transceivers to enable high-speed, reliable server-to-switch interconnection. In large-scale AI training and inference scenarios, they consistently deliver high-bandwidth, low-latency data transmission to fully unlock cluster performance.

Why NADDOD

Module Availability

Worry-Free - Rigorous authentic testing!
Multi-version Device Compatibility

Multi-Version - Easy adaptation for each module!
High-Performance

Actual Test Data-Better results intuitively!

Our Approach

Test scene

All tests were conducted in a high-density XDR environment to verify compatibility, connectivity, and long-term stability. NADDOD 800G DR4 transceivers on the NVIDIA HGX B300 server with mezzanine NVIDIA ConnectX-8 NICs were interconnected with NADDOD 1.6T DR8 transceivers on the fully loaded NVIDIA Quantum-X800 Q3400-RA switch, operating stably under continuous full-load conditions.

Test result

NVIDIA Quantum-X800 Q3400-RA Switch Version: 25.02.5002

NVIDIA HGX B300 (ConnectX-8 C8280U) Server Version: 40.46.5500

Contact us for more version testing

Multi-Version

Switch Side Testing
Server Side Testing

Multi-Version Testing

Verified across multiple software versions of the Quantum-X800 Q3400-RA switch, NADDOD 1.6T transceivers operate seamlessly, ensuring plug and play in real XDR clusters.
NADDOD 800G transceivers operate seamlessly across multiple software versions of the mezzanine ConnectX-8 NIC on the HGX B300 server, ensuring plug and play in real XDR clusters.
Connectivity

Switch Side Testing
Server Side Testing

Connectivity Testing

During continuous full-load operation, NADDOD 1.6T transceivers keep all switch ports up without link interruptions, providing stable, high-speed connectivity for high-density XDR clusters.
In the real deployment scenario, NADDOD 800G transceivers on the HGX B300 server, interconnected with NADDOD 1.6T transceivers on the switch, maintain stable end-to-end connectivity under continuous full-load conditions.
BER

Switch Side Testing
Server Side Testing

BER Testing

Under continuous full-load operation, NADDOD 1.6T transceivers achieve an ultra-low BER, ensuring consistent and reliable data transmission for demanding XDR workloads.
After long-term operation, NADDOD 800G transceivers on the HGX B300 server deliver an ultra-low BER, minimizing data retransmissions when interconnected with NADDOD 1.6T transceivers on the switch.

Questions & Answers

Ask a Question

Q:

What are the practical differences of 3nm DSP over common 5nm DSP solutions in real use?

A:

NADDOD's 800G XDR transceiver features Broadcom's 3nm DSP (BCM83628). It boasts lower power consumption and better overall performance than 5nm DSP solutions, while significantly lowering the port failure rate and further enhancing link stability in actual deployment. Tested on the NVIDIA ConnectX-8 C8180 network card, the 3nm solution delivers higher energy efficiency and stability, ensuring efficient and reliable operation of large AI clusters. For more details about this DSP technology and its advantages, see our in-depth blog: 200G/lane optical interconnect DSP technologies.

Q:

When should I choose XDR networking instead of NDR networking?

A:

For AI clusters with 512 to 1024 nodes, XDR networking is the best choice. At this scale, XDR networking delivers higher cost-efficiency and network reliability. Compared with NDR networking, it requires fewer switches and transceivers, thus reducing rack space, power consumption, simplifying cabling, and lowering O&M costs.

Q:

I am using NVIDIA DGX B300 or DGX GB300 servers now, can this 800G XDR transceiver be directly deployed?

A:

Yes. This 800G DR4 transceiver is designed for DGX B300/GB300 servers equipped with ConnectX-8 network cards. It connects to switches using 1.6T 2xDR4 transceivers, enabling high-speed links for large-scale AI clusters

Q:

I plan to buy about 2200 XDR transceivers in one batch for cluster upgrade, what is the current production capacity and expected delivery time?

A:

NADDOD's XDR transceivers have been deployed in batches in real AI clusters with proven mass production capability. We maintain strong supply capacity with short lead times, supporting large-scale project deployments. For specific procurement volumes and deployment plans, please contact your dedicated account manager or contact sales@naddod.com.

Q:

When building or upgrading an AI cluster, what application scenarios and computing scale require this 800G XDR transceiver?

A:

The 800G XDR transceiver is designed for AI computing scenarios with high network performance requirements, and are suitable for medium to ultra-large scale AI clusters. For medium and large-scale model training and inference scenarios, it meets the requirements of efficient data exchange and computing power expansion. For ultra-large GPU clusters and factory-level AI deployments, it delivers higher bandwidth density, ensuring smooth cluster operation and supporting stable long-term expansion.

Q:

Can the OSFP224-800G-DR4H module be used with QM9700/QM9790 series switches?

A:

No, this module is a single-channel 200G module with a flat-top form factor and cannot be used with the QM9700 series switches.

Q:

What is the difference between OSFP-800G-DR4H and OSFP-800G-2xDR4H?

A:

The OSFP-800G-DR4H is a single-port 800G transceiver based on 4x200G PAM4, typically used with ConnectX-8 network cards in GPU cluster architectures, connecting to 1.6T dual-port transceivers. In contrast, the OSFP-800G-2xDR4H is a dual-port transceiver based on 8x100G PAM4, mainly used for switch-to-switch interconnects such as NVIDIA Quantum-2 QM9700/9790. In short, OSFP-800G-DR4H is used for NIC-side connectivity in GPU cluster architectures, while OSFP-800G-2×DR4H is used for switch-side interconnects.

Q:

Can this transceiver be downgraded to 400G?

A:

Yes, when paired with a dual-port 1.6T transceiver OSFP-1.6T-2xDR4H and connected via a 12-fiber breakout cable, the transceiver can support XDR400 speed (dual-channel at 200G-PAM4 per channel).

Q:

What are the application scenarios for OSFP-800G-DR4H?

A:

The OSFP-800G-DR4H transceiver is primarily deployed on the NIC side in GPU cluster architectures, including ConnectX-8 and ConnectX-9 network cards. In NVIDIA DGX B300/GB300 architectures, it is typically used on the server side (DGX B300) and connects to OSFP-1.6T-2xDR4H transceivers on the switch side (Q3400-RA) via two 12-fiber MPO direct jumpers, enabling server-to-switch interconnect within the cluster.