DriveNets and AMD Publish Reference Architecture to Maximize AI Cluster Performance and Efficiency

Raanana, Israel – July 22, 2026 – DriveNets, a leader in large-scale networking solutions, today announced the publication of an end-to-end reference architecture, that defines a validated design blueprint for building high-performance AI infrastructure using AMD Instinct™ MI350 series GPUs and DriveNets AI Fabric. The combined solution supports scale-out and scale-across architectures, including front-end and storage networking based on a single networking solution. When combined with automated orchestration and full-stack integration services, it also supports rapid deployment and efficient end-to-end scaling.

The reference architecture and the companies’ growing strategic relationship reflect a shared vision for open, multi-vendor AI infrastructure. AMD recently participated as a strategic investor in DriveNets’ $410 million Series D financing round, reinforcing a strong commitment to their joint customers.

The reference architecture, available here, is accompanied by a complementary deployment guide that outlines a system-level approach for designing, deploying, and fine-tuning large AI GPU clusters. Together, the documents cover compute and networking design, as well as end-to-end optimization across the full-stack, including AMD ROCm™ software ecosystem, RCCL collective communications, network plugins, NICs, servers, and system-level orchestration to maximize training and inference performance.

Validated performances at scale

The reference architecture includes comprehensive benchmarking across inference, training, network isolation, and fabric resiliency workloads on AMD Instinct MI355X GPU clusters – providing validated, reproducible evidence of the platform’s readiness for production-scale AI deployments.

The architecture includes multiple network topology options, along with detailed technical guidance and best practices for optimizing performance, scalability, isolation, and resiliency.

Benchmark results show that DriveNets’ AI fabric delivers approximately 5% higher throughput and 10–15% lower time to first token (TTFT) compared to publicly available industry results. At multi-node scale, the platform meets strict production service-level objectives, including sub-20 millisecond inter-token latency and a minimum of 50 output tokens per second per user.

Resiliency testing demonstrated stable collective communication performance under concurrent RDMA traffic, with no degradation during transient link disruptions. Fabric bandwidth remained consistent, with no observable recovery delays. The results also demonstrate RCCL collective performance on AMD platforms comparable to leading publicly available NCCL benchmarks on equivalent configurations.

High performance with improved efficiency

These results validate the joint architecture as a scalable foundation for the full LLM lifecycle, enabling high utilization and consistent performance across training and inference workloads. The solution is designed to maximize GPU efficiency, reduce total cost of ownership, and deliver strong cost-per-token performance, offering an open alternative to vertically integrated AI infrastructure stacks.

“AI infrastructure is moving from single-vendor stacks to open, multi-vendor systems, and the network makes that shift work,” said Ido Susan, co-founder and CEO of DriveNets. “This reference architecture with AMD turns that into a validated blueprint: AMD Instinct GPUs and our AI fabric deliver higher throughput and lower time to first token than publicly available results — with the resiliency and consistency that production AI clusters require. This gives customers building on AMD Instinct a validated, end-to-end blueprint to maximize GPU utilization and lower cost-per-token, on open Ethernet.”

“Production AI clusters are defined by how well compute, network and software work together at scale,” said Arvind Balakumar, corporate vice president, Global Cluster Engineering, AMD. “Our collaboration with DriveNets gives customers a validated, open Ethernet reference architecture that pairs AMD Instinct MI350 series GPUs with DriveNets AI Fabric to help maximize GPU utilization, improve workload efficiency and simplify deployment across training and inference workloads.”

A better AI infrastructure choice

The deployment guide provides step-by-step instructions covering compute, network interface card (NIC), networking, collective-communications libraries and workload setup, configurations, and performance optimizations.

AMD and DriveNets are closely collaborating to ensure end-to-end software stack optimization, from NIC behavior and collective communications to system-level tuning. The companies are working with joint customers across production and proof-of-concept deployments, including LLM developers and NeoCloud providers. The companies also established a joint lab to enable customers to run proof-of-concept (PoC) tests using their own workloads.

About DriveNets

DriveNets is a leader in high-scale networking software for AI infrastructure and service providers. The company pioneered a disaggregated networking architecture that transforms the economics of large-scale networks while maximizing performance, utilization, and operational efficiency. DriveNets-powered networks are deployed by global leaders, including AT&T and Comcast, supporting more than 30% of total U.S. internet traffic. DriveNets AI Fabric delivers full-stack networking for AI infrastructures, providing the highest-performance, Ethernet-based alternative to InfiniBand. The solution is deployed by hyperscalers, Neo Clouds, and enterprises worldwide. Learn more at https://www.drivenets.com

About AMD

AMD drives innovation in high-performance and AI computing to solve the world’s most important challenges. Today, AMD technology powers billions of experiences across cloud and AI infrastructure, embedded systems, AI PCs and gaming. With a broad portfolio of AI-optimized CPUs, GPUs, networking and software, AMD delivers full-stack AI solutions that provide the performance and scalability needed for a new era of intelligent computing. Learn more at www.amd.com.

Frequently Asked Questions

What are the key performance benefits of the joint DriveNets and AMD reference architecture?

Benchmark results show that the joint architecture delivers approximately 5% higher throughput and 10–15% lower time to first token (TTFT) compared to publicly available industry results. At multi-node scale, the platform satisfies strict production service-level objectives, providing sub-20 millisecond inter-token latency and a minimum of 50 output tokens per second per user.

What components and technologies form the basis of this reference architecture?

The end-to-end blueprint integrates AMD Instinct™ MI350 series GPUs (benchmarked on MI355X clusters) with DriveNets AI Fabric. It covers full-stack optimizations across the AMD ROCm™ software ecosystem, RCCL collective communications, network plugins, NICs, servers, and system-level automated orchestration to support both scale-out and scale-across architectures.

How does the DriveNets and AMD solution handle network resiliency and traffic disruptions?

Resiliency testing demonstrated stable collective communication performance under concurrent RDMA traffic, with no performance degradation during transient link disruptions. The fabric bandwidth remained completely consistent with zero observable recovery delays, matching leading publicly available benchmarks on equivalent configurations.

{ “@context”: “https://schema.org”, “@graph”: [ { “@type”: “NewsArticle”, “@id”: “https://drivenets.com/news/drivenets-amd-reference-architecture-ai-performance/#article”, “headline”: “DriveNets and AMD Publish Reference Architecture to Maximize AI Cluster Performance and Efficiency”, “datePublished”: “2026-07-22T08:00:00+00:00”, “dateModified”: “2026-07-22T08:00:00+00:00”, “author”: { “@type”: “Organization”, “name”: “DriveNets”, “url”: “https://drivenets.com/” }, “publisher”: { “@type”: “Organization”, “name”: “DriveNets”, “logo”: { “@type”: “ImageObject”, “url”: “https://drivenets.com/logo.png” } } }, { “@type”: “FAQPage”, “@id”: “https://drivenets.com/news/drivenets-amd-reference-architecture-ai-performance/#faq”, “mainEntity”: [ { “@type”: “Question”, “name”: “What are the key performance benefits of the joint DriveNets and AMD reference architecture?”, “acceptedAnswer”: { “@type”: “Answer”, “text”: “Benchmark results show that the joint architecture delivers approximately 5% higher throughput and 10–15% lower time to first token (TTFT) compared to publicly available industry results. At multi-node scale, the platform satisfies strict production service-level objectives, providing sub-20 millisecond inter-token latency and a minimum of 50 output tokens per second per user.” } }, { “@type”: “Question”, “name”: “What components and technologies form the basis of this reference architecture?”, “acceptedAnswer”: { “@type”: “Answer”, “text”: “The end-to-end blueprint integrates AMD Instinct™ MI350 series GPUs (benchmarked on MI355X clusters) with DriveNets AI Fabric. It covers full-stack optimizations across the AMD ROCm™ software ecosystem, RCCL collective communications, network plugins, NICs, servers, and system-level automated orchestration to support both scale-out and scale-across architectures.” } }, { “@type”: “Question”, “name”: “How does the DriveNets and AMD solution handle network resiliency and traffic disruptions?”, “acceptedAnswer”: { “@type”: “Answer”, “text”: “Resiliency testing demonstrated stable collective communication performance under concurrent RDMA traffic, with no performance degradation during transient link disruptions. The fabric bandwidth remained completely consistent with zero observable recovery delays, matching leading publicly available benchmarks on equivalent configurations.” } } ] } ] }