Appreciation for the network has really grown in the past decade. Networks have played a crucial role in enabling digital transformation and, more recently, hybrid working, becoming the central nervous system of the modern enterprise. And the network is about to play its next starring role: an enabler of artificial intelligence.
AI is the newest frontier of data, and it’s the job of the network to get data to the right place, capturing, securing, transporting, and delivering data to wherever it’s needed. But with AI set to be the most data-heavy technology and compute-intensive workload of our time, are current enterprise networks ready to handle the demand?
Network readiness for AI
The answer is not yet, at least according to results from a January 2024 survey by Sapio Research commissioned by Hewlett Packard Enterprise. 1
Although 93% of the 2000-plus IT decision-makers we surveyed initially believed their network infrastructure was set up to support AI traffic, when we dug a little deeper, less than half said they fully understand the nuanced needs for networking across the full AI lifecycle, as shown in Figure 1. This may lead to inadequate provisioning.
IT decision-makers are actively investing in the network as part of their AI efforts. However, the network only ranked f ifth in their list of priority investment areas, coming in behind data management, security, software, and data protection. In overestimating the network’s readiness, IT leaders may not yet have given the network high enough consideration or investment it needs.
Understanding the network’s role in AI
There are many different stages in the AI lifecycle, from initial data gathering right through to inferencing and model retraining. The network plays a role in each stage and needs to be properly provisioned to:
Act as an on-ramp for building the data lake: For any AI system to make smart and accurate decisions, it first needs access to large and diverse datasets to learn from. The network must ensure data can move quickly and easily from any source to any destination — across edge, data center, and cloud. Additionally, the network needs to secure all the data it collects from the start, making sure it doesn’t get corrupted or tampered with, as this could poison the centralized data source a machine relies on to develop artificial intelligence.
Support model training and tuning: Once data has been captured and fed in, the next task is to train the algorithm. Training and retraining the models on huge amounts of information is an intense activity, requiring heavy horsepower matched with low-latency, high-performance network interconnects for hundreds or thousands of GPUs. When model training is done on-premises, the data center network needs to be up to the task of delivering optimized, power-efficient, and predictable performance.
Support inferencing: AI inferencing involves deploying newly trained AI models to where live data is created (such as the edge), so that the fresh data can be analyzed and acted upon quickly to generate business value. The network must operate from the edge to the cloud to continuously deliver a steady and coordinated flow of data to GPUs at the right time.
Protect AI: As organizations invest millions of dollars building extensive data lakes and high-performance training and inference solutions, both the data and the models need to be protected against data exfiltration, model theft, and rogue devices that can poison the data lake.
Three network imperatives to think about
To ensure the high-performing, resilient, and secure connectivity AI needs, here are three imperatives an enterprise needs to think strategically about:
1. Broad-based infrastructure: To satisfy the many different AI use cases, an enterprise needs to implement the broadest set of connectivity options — not just wired, Wi-Fi with an emphasis on IoT, and WAN but also private 5G — and be able to collect all data no matter where it’s generated.
2. Unified visibility and control: Implementing a broad set of connectivity options can easily lead to siloed infrastructure. Enterprises need a single connectivity fabric from edge to cloud, which will enable them to apply consistent security policies, centralize management, and understand the status of the network from a single pane of glass.
3. Security: Keeping bad actors away from data, applications, and infrastructure — and meeting compliance requirements — demands a network that has security built in, using Zero Trust and SASE principles.
Having flexibility is important too. AI is not static — it is a living solution, and the network must be able to respond to this.
An enabler and a beneficiary — AI enabling the network itself
While a dedicated network fabric is critical for enabling AI, there’s another side to the story: The network itself can also be enabled by AI. Enterprise networks are growing more diverse and distributed, to the point where managing and protecting them with traditional manual techniques is becoming almost impossible. Networks are only as good as the people managing them, and slips and blind spots are inevitable when operators are stretched and overwhelmed. This is where AI comes in.
AI for IT Operations (known as AIOps) combines big data, analytics, and machine learning to make life easier for networking teams. It offers 24/7 monitoring, automatically identifies issues before they impact the business, determines root causes, and delivers troubleshooting and optimization guidance in real time. Additionally, AI can support real-time endpoint profiling, so IT teams always know who and what is connecting to the network. Used this way, AI helps save time and money, improve security, and boost the efficiency of the team and network.
AI and the network
As you can see, modern AI applications need to be supported by high-bandwidth, low-latency, secure, and scalable networks. And these networks must operate across the AI lifecycle from data capture and aggregation to high-performance model training and edge-to-cloud inference deployment. Similarly, modern networks need to be supported by AI. Safe to say, the two have an important future together, where one cannot perform at its best (or at all) without the other.
1. “Architect an AI advantage,” Hewlett Packard Enterprise, 2024
This article originally appeared in the June 2024 edition of The Doppler
Related resources:
Sylvia_Hooks
VP, Edge to Cloud Integrated Marketing, Hewlett Packard Enterprise