How MMC Connectors Support AI-driven Data Center Architecture

As AI models scale exponentially, AI data centers demand higher bandwidth to support massive data transfers, lower network latency to ensure real-time responsiveness, and more compact cabling designs to accommodate device-dense deployments. The MMC connector, as the latest innovation in the VSFF (Very Small Form Factor) fiber optic connector, meets these evolving needs by offering a compact form factor and reduced insertion loss.

Fiber Cabling Trends in AI

Trend Towards High-Capacity, High-Density Clusterization

AI applications, especially in deep learning and big data analytics, require massive computational and storage resources. Therefore, AI data center cabling is shifting towards high-capacity, high-density clusterization. The NVIDIA DGX SuperPOD AI data center infrastructure platform is a typical example of this trend. It consists of multiple DGX systems, GPUs, storage, and high-performance networking equipment, all interconnected by high-density fiber optic cabling to ensure fast data flow between nodes.

Within its InfiniBand (IB) computing network, the Leaf-Spine links require multiple inter-cabinet and inter-row cable development. The total number of cables reaches 5,104 (as shown in Table 1), and this is an idealized layout—actual construction requires an even greater number of cables. This highlights the enormous fiber optic connectivity required in AI data centers to tightly network dozens or even hundreds of GPUs into 400G/800G clusters, facilitating distributed training and inference for AI models.

			InfiniBand Switch Count		Cable Counts
SU Count	Node Count	GPU Count	Leaf	Spine	Compute and UFM	Spine-Leaf
1	31	248	8	4	252	256
2	63	504	8	8	508	512
3	95	760	16	16	764	768
4	127	1016	16	16	1020	1024

Table 1: Numbers of fiber cables required in IB network (figures come from NVIDIA DGX SuperPOD AI infrastructure).

Shift Towards Higher Quality Fiber Components

AI training and inference are highly sensitive to link latency, especially when processing real-time computations such as financial transaction analysis or smart manufacturing control. Even the slightest fluctuation in optical signals can cause system delays or errors. Therefore, fiber optic connectors must have lower insertion loss and reflection loss to minimize signal attenuation and maintain transmission quality.

In this context, Angled Physical Contact (APC) connectors offer significant advantages over Ultra Physical Contact (UPC) connectors. APC connectors, with their unique angled polishing method, can significantly reduce back-reflections, provide superior return loss characteristics and ensure more stable and reliable performance. At data rates of 100G and above, especially when using PAM4 signaling for higher bandwidth, the reduction of back-reflection becomes even more crucial. According to training materials from CommScope, APC connectors demonstrate significantly better return loss performance in the 50G PAM4 channel compared to UPC connectors. This is crucial for maintaining real-time data accuracy and system efficiency in AI-driven environments.

Figure 1: APC and UPC Performance in 50G PAM4 Channel.

Practical Applications and Advantages of MMC Connectors in AI Data Centers

Simplifying Cabling for AI Model Training

As previously mentioned, in the case of the NVIDIA DGX SuperPOD H100, the deployment of the InfiniBand (IB) computing network alone requires a staggering 5,104 fiber optic cable connections. Completing such an extensive fiber connection will take several weeks, highlighting the need for more efficient cabling solutions.

The MMC connector offers a clear advantage with its compact TMT sleeve design. It provides three times the port density of MTP/MPO connectors in a 1U rack when comparing similar fiber counts (e.g., MMC-16 fiber vs. MTP-16 fiber). This reduces the space required for fiber connections and allows more ports to be packed into the same rack or cabinet.

Additionally, the MMC's DirectConec™ push-pull tail design enables faster installations and eliminates the need for extra finger space during insertion and removal, maximizing available panel space. It increases rack density and leaves more room for airflow, which helps mitigate potential heat dissipation issues caused by cable congestion. This can improve the overall thermal management of the data center, ensuring that the AI infrastructure runs smoothly and efficiently. Learn more MMC connectors in high-density cabling, please click the article How MMC Connectors Optimize High-Density Fiber Networks.

Supporting Real-time AI Inference in Edge Computing Environments

Based on official data from US Conec, the MMC connector's performance in the GR-1435 test shows that its optimized TMT insert design achieves insertion loss lower than the IEC Grade B standard (see Insertion Loss Test) and excellent end-face flatness (see End-face Flatness Test). This effectively reduces optical signal loss and ensures the stability of high-speed transmissions at 400G, 800G, and 1.6T. Additionally, in harsh environmental tests such as thermal aging and humidity aging, the insertion loss variation is less than 0.1 dB (see Environmental Test), confirming its long-term reliability in complex environments. Therefore, the MMC connector can provide efficient and stable fiber links for edge AI devices, accelerating data transmission and processing to support the efficient operation of real-time AI inference.

Figure 2: MMC connector's performance tests.

Adapting to Next-Generation Network Architectures

As mentioned above, the MMC connector uses an optimized TMT insert design, offering three times the port density of MTP/MPO in the same space. At the same time, the TMT sleeve retains the alignment structure of the traditional MT sleeve. By using hybrid adapters, the MMC connector can seamlessly connect with MTP/MPO connectors. This makes it particularly suitable for MTP/MPO fiber deployment scenarios that require a gradual transition to high-density, high-speed AI architectures with MMC deployment. In next-generation network architectures that demand extremely low latency (such as financial computing, edge AI inference, and other scenarios), the MMC connector can also provide stable fiber links, perfectly aligning with the development direction of next-generation network architectures.

Furthermore, the MMC ecosystem continues to mature with the growing availability of multi-vendor components and terminations. FS has already introduced 16-core MMC fiber jumpers APC to support the deployment of 800G networks. Looking ahead, FS will continue to invest in research and development to optimize the performance of MMC products, working alongside the industry to meet the ever-expanding possibilities of the AI and high-speed networking era.