AI Networking Fabric

AI infrastructure spending shift from training to inference is driving the adoption of Heterogeneous AI architectures that bring infrastructure costs down and optimize power utilization.

Heterogeneous AI architecture uses multiple AI accelerators from multiple vendors within the same cluster, each best for a different stage or task within the AI training or inference process.

The Heterogeneous AI cluster is orchestrated to provide the best performance and power utilization, to reduce the overall cost and power usage.

DriveNets’ AI fabric can uniquely support Heterogeneous AI, with:

• Full-stack performance optimization for any AI accelerator and vendor
• Robust networking infrastructure enabling the highest performance for the bursty and unpredictable traffic patterns of heterogeneous AI
• Open eco-system, use any GPU, NIC and storage

Scale-across networking is an extension of the backend network to a remote site/datacenter. Unlike DCI (datacenter interconnect) which extends frontend networks, scale-across requires the highest quality of service in terms of packet-loss, tail-latency and jitter.

Large language models (LLMs), generative AI, and advanced analytics are pushing data center infrastructures to their limits. Enterprises in industries like automotive, finance and pharmaceuticals, as well as GPU-as-a-service (GPUaaS or NeoClouds) providers, are building their own AI infrastructure and scaling beyond a single data center to meet surging AI demand. This expansion addresses the growing computational demands, power and space limitations, resiliency needs, and data locality requirements of modern AI workloads.

To connect AI clusters across multiple remote locations into a super-cluster, DriveNets enhance its intra-site scheduled-fabric solution with “interconnect” NCPs (leaf white boxes) powered by the Jericho3-AI variant with high-bandwidth memory (HBM) buffers. These deep buffers absorb bursts of AI traffic that would otherwise lead to congestion and packet loss.

AI hyperscaler networks demand unparalleled scalability, performance, and simplicity to support the needs of very-large-scale GPU clusters. DriveNets AI Fabric rises to this challenge with a software-driven architecture that supports extremely large GPU cluster sizes, leveraging Ethernet technology to ensure seamless integration and operation. Unlike traditional solutions, DriveNets eliminates the need for specialized knowledge from technical staff, simplifying deployment and management. It’s easy-to-scale, distributed, disaggregated design enables hyperscalers to expand capacity effortlessly while maintaining optimal performance and efficiency—empowering innovation without the limitations of proprietary or complex networking systems.

Best performance solution for Hyperscalers AI-fabric

• Better performance than InfiniBand
• Native support for multi-tenancy
• Multi-site cluster without performance compromise
• Unified, Ethernet based, fabric for storage and compute
• Field proven – in large hyperscaler production

AI enterprise backend networks demand exceptional performance and scalability to support the growing reliance on AI-driven workloads and data-intensive applications. DriveNets AI Fabric addresses these needs with a distributed, scheduled fabric architecture—scalable, cost-effective, and optimized for AI workloads that fits all Enterprise use cases including financial research, life science and pharmaceutical research, automotive, energy & utilities and high education. Its Fabric Scheduled Ethernet and Endpoint Scheduled Ethernet technologies enable enterprises to seamlessly adapt to evolving AI demands, ensuring high performance, no vendor lock, operational simplicity, and the ability to innovate without the constraints of proprietary network designs.

• Best Ethernet based AI-fabric
• Better performance than InfiniBand
• Fastest time to deploy AI clusters
• Any Optics, any NIC, any GPU