Blog: 3 reasons why nanosecond-level synchronisation is essential for handling AI workloads in data centres

As artificial intelligence (AI) and high-performance computing (HPC) applications continue to expand, synchronisation in data centres has become increasingly critical. Modern AI workloads demand precise timing to ensure seamless parallel processing, real-time data handling, and efficient resource use. This article explores why synchronisation is essential for AI deployment in data centres, focusing on its impact on parallel workloads, time-sensitive processing, and enhanced user experiences.

Today’s traditional data centres already require synchronisation.

Even before the rise of AI, synchronisation in data centres was already playing an important role in database management and content delivery to ensure consistency, reliability, and optimal performance. Let's examine this in more detail:

Database Synchronisation: Data centres operate vast distributed databases that must remain consistent across multiple servers and geographical locations. Synchronisation is crucial for:

• Transaction Coordination: In distributed databases, transactions must be executed consistently to prevent data corruption. Timestamp-based synchronisation helps ensure that transactions across different nodes are correctly ordered.
• Commit-Wait Reduction: In high-frequency transaction systems, commit-wait cycles—where one transaction must wait for another to complete—can introduce delays. Precise synchronisation minimises these delays, enhancing performance.
• Replication Consistency: Many databases use replication mechanisms to maintain backups and distribute workloads. Time synchronisation ensures that replicated data remains identical across nodes, preventing conflicts or outdated information.

Optimised Search and Query Performance: Search engines and large-scale data indexing systems rely on synchronised caches to ensure efficient information retrieval:

• Cache Invalidation: A sub-microsecond difference in time synchronisation can result in serving outdated data from the cache. Ensuring precise time alignment helps invalidate stale cache entries and maintain consistency.
• Load Balancing Accuracy: Queries are distributed across multiple nodes in search-intensive environments. Accurate synchronisation prevents query duplications or mismatches, improving search speed and reliability.

Real-Time Content Delivery and Event Ordering: Cloud services and content delivery networks (CDNs) depend on synchronisation to manage large-scale media distribution and user interactions.

• Precise Timestamping: Content updates, user interactions, and data modifications must be accurately timestamped to ensure event ordering. Without precise time synchronisation, users may experience inconsistencies such as seeing outdated content or incorrect playback sequences.
• Live Streaming and Online Media: Streaming services require precise synchronisation between content servers to prevent buffering, latency, or mismatched audio/video feeds.
• E-Commerce and Financial Transactions: Online platforms processing thousands of transactions per second need event ordering to prevent duplicate payments, transaction fraud, or discrepancies in purchase logs.

Reduced Response Time for End Users: In cloud computing environments, synchronisation plays a key role in improving application responsiveness:

• Load Distribution: Cloud applications distribute workloads across multiple data centres. Accurate time synchronisation ensures that data processing is evenly distributed, preventing the overloading of specific servers.
• Predictable Latency: Users expect low variance in response times, particularly in interactive applications such as online gaming, financial trading, and remote collaboration tools. Synchronisation helps maintain consistent and predictable response times.

Inadequate synchronisation is letting down AI workloads today

AI workloads are pushing synchronisation requirements to a whole new level—down to the nanosecond.

AI workloads, particularly training models, often involve synchronised parallel jobs across multiple nodes. According to research from Meta, up to 33% of AI elapsed time is spent waiting for the network. This highlights the inefficiencies that arise when time synchronisation is inadequate.

For effective AI model training, multiple GPUs or TPUs must operate in tandem, exchanging data continuously. If these computations are not perfectly aligned in time, discrepancies arise, leading to increased latency and bottlenecks. Synchronisation ensures that all computational nodes process data with minimal time delays, reducing idle times and improving throughput.

1. Importance of Time and Order-Sensitive Processing: Many AI applications require precise time alignment to ensure consistency and accuracy. This is particularly critical in:

• Frame-accurate correlation: AI-driven video analytics require synchronisation across distributed encoders and decoders to maintain coherence.
• Audio-visual synchronisation: AI in media processing must perfectly align audio and video streams to prevent lip-sync errors and distortions.
• Parallel video processing pipelines: AI entertainment, surveillance, and medical imaging applications demand synchronised time stamps to maintain real-time accuracy.

Without effective synchronisation, errors in frame sequencing and processing delays can lead to degraded output quality and unreliable AI inference.

2. Real-Time Processing for AI Applications: Synchronisation is also a key enabler for real-time AI processing in critical applications such as:

• Autonomous vehicles: AI-powered self-driving systems require millisecond-level precision to process sensor data, make split-second decisions, and ensure safety.
• Live video analytics: Surveillance systems depend on synchronised video feeds to detect anomalies and track movements accurately.
• Industrial automation: AI-driven robotics in manufacturing rely on synchronised control loops to execute precise movements without latency issues.

Even minor synchronisation errors can lead to catastrophic failures in these scenarios, making precise timing a non-negotiable requirement.

3. Enhancing the User Experience in 3D Simulations: AI is increasingly used in virtual reality (VR) and augmented reality (AR) applications, where precise video timing across multiple nodes is crucial for a seamless experience. Delays or desynchronisation in rendering different frames can result in motion blur, lag, or discomfort for users.

By implementing sub-microsecond synchronisation, data centres can ensure that all nodes in a 3D simulation render images at precisely the right moment, reducing visual artefacts and improving overall realism.

How can you achieve 10s of ns synchronisation across AI server systems?

AI data centres rely on synchronised physical layers to ensure all nodes operate within a unified clock domain. This is achieved by embedding precise timestamping at the PHY layer, where every node retrieves synchronisation signals with nanosecond-level accuracy.

To achieve overall synchronisation across all nodes in an AI server system to within tens of nanoseconds, data centre architects need to consider the following:

• Recovered Clock Mechanisms: synchronise the end clients with the AI cluster’s master clock using digitally controlled oscillators, ensuring a stable and traceable time source.
• Low Jitter Clocks: use the stability and phase noise of low jitter reference clocks on the transceivers to enable high data throughput.
• Precision Timing Protocols (PTP): use PTP-enabled network interface modules to efficiently synchronise time-sensitive workloads of the AI servers.
• Optimised Servo Loops: use optimised servo loops to ensure minimal phase deviation and reduce clock drift to maintain synchronisation across the AI processing pipeline.

By leveraging these advanced clocking techniques, AI server systems can achieve ultra-low latency, optimised throughput, and consistent processing cycles across all workloads. To learn more about Rakon’s high-precision oscillators for AI workloads, please visit https://www.rakon.com/datacentres 

Would you like more content like this delivered to your inbox?

Subscribe to our emails now!