How hole core fiber is accelerating AI  

One among these applied sciences that was highlighted at Microsoft Ignite this previous November was Hole Core Fiber, an progressive optical fiber that’s set to optimize the Microsoft Azure world cloud infrastructure, providing superior community high quality and safe knowledge transmission.

This weblog is a part of the ‘Infrastructure for the period of AI’ sequence that focuses on rising know-how and traits in large-scale computing. This piece dives deeper into considered one of our latest applied sciences, hole core fiber (HCF). 

AI is on the forefront of individuals’s minds, and improvements are taking place at lightning pace. However to proceed the tempo of AI innovation, firms want the appropriate infrastructure for the compute-intensive AI workloads they’re making an attempt to run. That is what we name ‘purpose-built infrastructure’ for AI, and it’s a dedication Microsoft has made to its clients. This dedication doesn’t simply imply taking {hardware} that was developed by companions and putting it in its’ datacenters; Microsoft is devoted to working with companions, and sometimes by itself, to develop the latest and best know-how to energy scientific breakthroughs and AI options. 

One among these applied sciences that was highlighted at Microsoft Ignite in November was hole core fiber (HCF), an progressive optical fiber that’s set to optimize Microsoft Azure’s world cloud infrastructure, providing superior community high quality, improved latency and safe knowledge transmission. 

Transmission by air 

HCF know-how was developed to fulfill the heavy calls for of workloads like AI and enhance world latency and connectivity. It makes use of a proprietary design the place mild propagates in an air core, which has vital benefits over conventional fiber constructed with a stable core of glass. An attention-grabbing piece right here is that the HCF construction has nested tubes which assist cut back any undesirable mild leakage and hold the sunshine moving into a straight path by the core.  

As mild travels quicker by air than glass, HCF is 47% quicker than normal silica glass, delivering elevated total pace and decrease latency. It additionally has the next bandwidth per fiber, however what’s the distinction between pace, latency and bandwidth? Whereas pace is how shortly knowledge travels over the fiber medium, community latency is the period of time it takes for knowledge to journey between two finish factors throughout the community. The decrease the latency, the quicker the response time. Moreover, bandwidth is the quantity of information that’s despatched and obtained within the community. Think about there are two autos travelling from level A to level B setting off on the identical time. The primary automobile is a automobile (representing single mode fiber (SMF)) and the second is a van (HCF). Each autos are carrying passengers (which is the information); the automobile can take 4 passengers, whereas the van can take 16. The autos can attain completely different speeds, with the van travelling quicker than the automobile. This implies it would take the van much less time to journey to level B, due to this fact arriving at its vacation spot first (demonstrating decrease latency).  

For over half a century, the trade has been devoted to creating regular, but small, developments in silica fiber know-how. Regardless of the progress, the features have been modest as a result of limitations of silica loss. A major milestone with HCF know-how was reached in early 2024, attaining the bottom optical fiber loss (attenuation) ever recorded at a 1550nm wavelength, even decrease than pure silica core single mode fiber (SMF). ¹ Together with low attenuation, HCF gives larger launch energy dealing with, broader spectral bandwidth, and improved sign integrity and knowledge safety in comparison with SMF. 

The necessity for pace 

Think about you’re enjoying a web based online game. The sport requires fast reactions and split-second selections. In case you have a high-speed reference to low latency, your actions within the sport will probably be transmitted shortly to the sport server and to your pals, permitting you to react in actual time and luxuriate in a clean gaming expertise. Alternatively, when you have a gradual reference to excessive latency, there will probably be a delay between your actions and what occurs within the sport, making it troublesome to maintain up with the fast-paced gameplay. Whether or not you’re lacking key motion instances or lagging behind others, lagging is extremely annoying and might critically disrupt gameplay. Equally, in AI fashions, having decrease latency and high-speed connections will help the fashions course of knowledge and make selections quicker, enhancing their efficiency. 

Decreasing latency for AI workloads

So how can HCF assist the efficiency of AI infrastructure? AI workloads are duties that contain processing giant quantities of information utilizing machine studying algorithms and neural networks. These duties can vary from picture recognition, pure language processing, laptop imaginative and prescient, speech synthesis, and extra. AI workloads require quick networking and low latency as a result of they typically contain a number of steps of information processing, resembling knowledge ingestion, preprocessing, coaching, inference, and analysis. Every step can contain sending and receiving knowledge from completely different sources, resembling cloud servers, edge units, or different nodes in a distributed system. The pace and high quality of the community connection have an effect on how shortly and precisely the information might be transferred and processed. If the community is gradual or unreliable, it could actually trigger delays, errors, or failures within the AI workflow. This may end up in poor efficiency, wasted assets, or inaccurate outcomes. These fashions typically want big quantities of processing energy and ultra-fast networking and storage to deal with more and more refined workloads with billions of parameters, so finally low latency and high-speed networking will help pace up mannequin coaching and inference, enhance efficiency and accuracy, and foster AI innovation. 

Serving to AI workloads in every single place

Quick networking and low latency are particularly essential for AI workloads that require real-time or near-real-time responses, resembling autonomous autos, video streaming, on-line gaming, or good units. These workloads have to course of knowledge and make selections in milliseconds or seconds, which implies they can’t afford any lag or interruption within the community. Low latency and high-speed connections assist make sure that the information is delivered and processed in time, permitting the AI fashions to offer well timed and correct outcomes. Autonomous autos exemplify AI’s real-world software, counting on AI fashions to swiftly determine objects, predict actions, and plan routes amid unpredictable environment. Speedy knowledge processing and transmission, facilitated by low latency and high-speed connections, allow close to real-time decision-making, enhancing security and efficiency. HCF know-how can speed up AI efficiency, offering quicker, extra dependable, and safer networking for AI fashions and purposes. 

Regional implications 

Past the direct {hardware} that runs your AI fashions, there are extra implications. Datacenter areas are costly, and each the space between areas, and between areas and the shopper, make a world of distinction to each the shopper and Azure because it decides the place to construct these datacenters. When a area is positioned too removed from a buyer, it leads to larger latency as a result of the mannequin is ready for the information to go to and from a middle that’s additional away.

If we take into consideration the automobile versus van instance and the way that pertains to a community, with the mix of upper bandwidth and quicker transmission pace, extra knowledge might be transmitted between two factors in a community, in two thirds of the time. Alternatively, HCF gives longer attain by extending the transmission distance in an current community by as much as 1.5x with no affect on community efficiency. In the end, you may go an additional distance on the identical latency envelope as conventional SMF and with extra knowledge. This has big implications for Azure clients, minimizing the necessity for datacenter proximity with out growing latency and decreasing efficiency. 

The infrastructure for the period of AI 

HCF know-how was developed to enhance Azure’s world connectivity and meet the calls for of AI and future workloads. It gives a number of advantages to finish customers, together with larger bandwidth, improved sign integrity, and elevated safety. Within the context of AI infrastructure, HCF know-how can allow quick, dependable, and safe networking, serving to to enhance the efficiency of AI workloads. 

As AI continues to evolve, infrastructure know-how stays a crucial piece of the puzzle, making certain environment friendly and safe connectivity for the digital period. As AI developments proceed to position extra pressure on current infrastructure, AI customers are more and more in search of to profit from new applied sciences like HCF, digital machines just like the just lately introduced ND H100 v5, and silicon like Azure’s personal first associate AI accelerator, Azure Maia 100. These developments collectively allow extra environment friendly processing, quicker knowledge switch, and finally, extra highly effective and responsive AI purposes. 

Sustain on our “Infrastructure for the Period of AI” sequence to get a greater understanding of those new applied sciences, why we’re investing the place we’re, what these developments imply for you, and the way they permit AI workloads.   

Extra from the sequence

Sources

¹ Hole Core DNANF Optical Fiber with <0.11 dB/km Loss