When understood, these six concepts will help anyone new to ACI to understand a more detailed technical discussion.

Widely available GPU-accelerated servers, combined with better hardware and popular programming languages like Python and C/C++, along with frameworks such as PyTorch, TensorFlow and JAX, simplify the development of GPU-accelerated ML applications. These applications serve diverse purposes, from medical research to self-driving vehicles, relying on large datasets and GPU clusters for training deep neural networks. Inference frameworks apply knowledge from trained models to new data with optimized clusters for performance. The learning cycles involved in AI workloads can take days or weeks, and high-latency communication between server clusters can significantly impact completion times or result in failure. AI workloads demand low-latency, lossless networks, requiring appropriate hardware, software features, and configurations. This article will explain advanced queueing solutions used by all the major OEMs in the Network Operating Systems (NOS) that support ECN and PFC.

RoCEv2 is a solution for achieving swift data throughput and minimal delay in modern data centers. It incorporates features like Priority Flow Control (PFC) and Explicit Congestion Notification (ECN) to establish a lossless network environment. PFC manages data flow at the interface level, while ECN detects and mitigates congestion before PFC activation becomes necessary. The combination of ECN and PFC, known as Data Center Quantized Congestion Notification (DCQCN), optimizes congestion management in RDMA networks. Careful tuning of queue thresholds is crucial to prevent hot spots and ensure low Job Completion Times (JCTs). The use of ECN and PFC is necessary for maintaining a lossless fabric in GPU-to-GPU communication during AI/ML training runs.

AI workloads require significant data and computational power, with billions of parameters and complex matrix operations. Inter-network communication accounts for a significant portion of job completion time. Traditional network architectures are insufficient for large-scale AI training, necessitating investments in new network designs. Arista Networks offers high-bandwidth, low-latency and scalable connectivity for GPU servers, with features like Data Center Quantized Congestion Notification and intelligent load balancing. Arista's AI Leafs and Spines provide high-density and high-performance switches for AI networking. Different network designs are recommended based on the size of the AI application. A dedicated storage network is recommended to handle the large datasets used in AI training. Arista's Cloud Vision Portal and AI Analyzer tools provide automated provisioning and deep flow analysis. Arista's IP/Ethernet switches are well-suited for AI/ML workloads, offering energy-efficient interconnects and simplified network management.

This article discusses the importance of deploying AI applications and training models using distributed computing and the need for significant computational resources. It highlights the role of network efficiency and scalability in large-scale AI deployments.

An in-depth guide for IT decision makers navigating the complex IT landscape.

This article examines what End of Support/End of Life means, how it can affect your business and steps to making a plan to refresh your data center.

Learn how to create a 90-day POC to verify fabric performance, troubleshoot issues and validate the usefulness of the NEXUS Dashboard and day 2 operations suite.

ACI is a powerful technology offering rich features for SDN to include application-centric security segmentation, automation and orchestration in the data center.

The second in our series, this article explores what intent-based networking (IBN) is and how organizations can leverage it to build multi-domain architectures.

This article is part of our WWT Primer Series, a collection of content focused on the fundamental understanding of complex technologies and solutions.

Optical Data Center Interconnect (DCI) provides a cost saving, high density, flexible alternative to leased circuits for connecting geographically separated data centers.