Siemon is a customer of Business Reporter.
The most recent advancements in artificial intelligence/machine learning and the necessary resources for your physical IT infrastructure to handle these emerging requirements.
In the rapidly evolving realm of AI technology, new developments surface nearly every day. Clearly, AI possesses a significant capacity to transform our lives, a technology that spans chatbots, facial recognition, self-driving vehicles, and early disease detection.
In 2023, the worldwide market for artificial intelligence (AI) was worth $142.3 billion, with industries such as finance, healthcare, and high-tech/telecommunications leading the way in implementing AI.
Artificial intelligence is currently employed for monitoring assets in data centers, identifying faults before they occur, and optimizing energy consumption to improve power usage effectiveness (PUE). This technology is not only utilized by Hyperscalers, but also by numerous large enterprise companies.
InfiniBand versus Ethernet
Ethernet continues to dominate as the primary global standard in the majority of data centers. However, there is a growing trend in modern AI networks to utilize InfiniBand technology. Currently, InfiniBand holds a small portion of the market share, mainly being utilized in HPC networks.
There is now a competition developing among leaders in the InfiniBand market and well-known manufacturers of Ethernet switches and chips. These manufacturers have created new chips that can be used to build AI clusters using Ethernet instead of InfiniBand. Both InfiniBand and Ethernet require high bandwidth and low latency, so the use of high-quality optical cables is essential for optimal performance in large language model (LLM) training and inferencing, regardless of the chosen protocol.
Exponential demands for power and bandwidth
Two major difficulties that data centers are currently encountering involve the high energy demands and cooling demands for equipment, as well as the excessive bandwidth requirements for GPUs.
Supercomputers with GPUs running AI applications demand vast power and multiple high-bandwidth connections. These GPUs demand from 6.5kW to over 11kW per 6U unit. When contrasted with packed data center cabinets, averaging 7-8kW and maxing at 15-20kW per cabinet, the extent of AI’s power appetite becomes clear. Many of the leading Server OEMs are also offering servers with these GPUs.
These GPUs typically need connections with bandwidth of up to 8x100Gb/s (EDR), 200Gb/s (HDR) or 400Gb/s (NDR). Every node commonly has eight connections, equating up to 8x400G or 3.2 terabit per node.
How will the IT infrastructure handle these requirements?
The increasing power and cooling needs of data centers are causing network managers to reassess their infrastructure. This may include adjusting network designs and spacing out GPU cabinets more, potentially utilizing end-of-row (EoR) setups to effectively manage rising temperatures.
This results in a larger distance between switches and GPUs. To address this, data centers may need to use more fiber cables for switch-to-switch connections. Due to the longer distances, direct attach cables (DACs) may not be suitable as they are limited to only five meters at these speeds.
AOCs, or active optical cables, are a viable option due to their ability to cover longer distances than DACs. They also have the benefit of lower power consumption and improved latency compared to transceivers. Siemon offers AOCs in 0.5m increments, making cable management easier.
Upgrading the connections between switches in a data center requires the use of parallel optic technology to handle the growing need for higher bandwidth. There are currently options available that utilize eight fibers and MPO/MTP connectors for parallel fiber optic technology. These MPO Base-8 solutions allow for the use of either singlemode or multimode fibers and make it easier to upgrade to faster speeds. For enterprise data centers, it is recommended to consider a Base-8 MPO OM4 cabling solution when moving to 100Gb/s and 400Gb/s speeds. On the other hand, cloud data centers should opt for a Base-8 MPO singlemode cabling solution when transitioning to 400Gb/s and 800Gb/s speeds.
New fiber enclosures available on the market offer versatile support for various fiber modules, such as Base-8 and Base-12 with LC shutters, MTP pass-thru modules, and splicing modules. These enclosures provide convenient accessibility and enhanced cable organization.
In the world of AI technology, speed is extremely important. Siemon recommends using “AI-Ready” options that use ultra-low loss (ULL) performance and MTP/APC connectors. Ultra-low-loss fiber connections are crucial for new short-distance singlemode applications, which support speeds of 100, 200, and 400 Gb/s over distances longer than 100 meters. This type of connectivity meets the strict requirements for insertion loss set by AI applications, improving overall network performance.
Siemon also recommends using APC (angled physical connect) fiber connectors, such as the MTP/APC type, for certain multimode cabling uses, in addition to the standard singlemode method. The angled end-face design of APC connectors (compared to UPC connectors) decreases reflectance, resulting in improved fiber performance.
AI is a revolutionary technology with the potential to completely change our work and personal lives. As data centers prepare for its impact, they should prioritize implementing measures to accommodate faster data speeds and improve energy efficiency. By effectively preparing for AI, data center operators will be able to take advantage of the opportunities that come with its ongoing development and widespread use.
To learn more, go to siemon.com/ai.
Source: independent.co.uk
