Cornelis Announces New Reference Architecture for AI Inference, Training, and HPC Built for AMD 6th Gen EPYC™ and AMD Instinct™ MI400 Series

CN6000-based design helps operators scale and improve time to outcome for agentic AI and HPC, builds on a five-year collaboration between Cornelis and AMD.

WAYNE, Pa.--(BUSINESS WIRE)--Cornelis, a leading provider of high-performance networking solutions, unveiled a new reference architecture that pairs its CN6000 SuperNIC with AMD 6^th Gen EPYC™ processors and AMD Instinct™ MI400 series GPUs. Designed for disaggregated AI inference, large-scale training, and HPC simulation, the architecture provides infrastructure operators with a blueprint for building high-performance systems on an integrated compute and networking platform.

Cornelis plans to detail the full architecture and share OEM validation results this fall at OCP Global Summit and the AI Infra Summit.

The new reference architecture comes as AMD prepares to detail its next generation of AI infrastructure during the Advancing AI 2026 event, taking place July 22-23 in San Francisco. The announcement reflects Cornelis’ continued investment in enabling customers deploying the latest generation of AMD compute platforms.

Engineering High-Performance AI Networks

AI infrastructure now runs multiple workloads that share the same dependency: a network that keeps pace with the compute sitting next to it. Training synchronizes updates across thousands of accelerators at once. Inference has split into prefill, decode, and agent orchestration stages that hand work to each other in sequence. HPC simulation moves large volumes of state between nodes on every iteration. In each case, a network that queues a transfer makes the compute behind it wait.

AMD's latest roadmap reflects that shift. Its EPYC™ Venice processor, built on a 2-nanometer process with up to 256 cores per socket, is expected to become the industry’s first CPU to reach that milestone on TSMC's most advanced node. AMD has also previously detailed its Instinct™ MI400 series GPUs, for large-scale cloud training and inference, on-premises enterprise deployments, and sovereign AI and HPC.

Cornelis designed the CN6000 for the density that these powerful new compute offerings create. The SuperNIC supports Omni-Path, RoCEv2, and Ultra Ethernet transport on a single adapter, so operators can standardize their network strategy without locking into one protocol. It is engineered for 800 Gbps of bandwidth through a native PCIe 6.0 x16 host interface, matching AMD’s latest offerings, so the network and the compute advance together. The Cornelis CN6000 architecture is targeting more than 1.6 billion bidirectional messages per second. This keeps time to first token low and token throughput steady as concurrent requests scale.

Cornelis has begun modeling the efficiency and ROI potential of this new architecture, starting with AI training, where pre-production simulations show a CN6000 based network completing AllReduce collectives roughly 24 percent faster than standard Ethernet. This evaluation also resulted in cutting training time for a 250-billion-parameter model on a simulated 10,000 GPU cluster by about 13 percent. The same collective-communication advantage extends directly to inferencing, where disaggregated architectures depend on fast, low-latency exchange between prefill and decode nodes and MoE expert routing, exactly the traffic pattern CN6000 is built to accelerate. The CN6000 family pairs the 800G multi-protocol SuperNIC with a 48-port Omni-Path switch and a libfabric-based software stack that is seamlessly compatible with AMD ROCm™. The CN6000 is expected to be generally available in Q4CY26.

Building on a Proven Collaboration

The architecture extends Cornelis’s collaboration in market with AMD that spans over five years, more than 25 validated platform configurations, and nine OEM partners across the two companies' shared ecosystem. The two companies also operate a joint HPC Center of Excellence in Munich, where engineering teams from both sides validate performance on real workloads ahead of customer deployment. The reference architecture outlined today builds on that foundation and points it toward AMD's newest CPU and GPU generation.

“AMD built its CPU and AI accelerator business by giving operators a real alternative when the market didn't offer many,” said Lisa Spelman, chief executive officer of Cornelis. “We built Cornelis the same way. As AMD gets ready to share more of its AI vision and roadmap, Cornelis is proud to be one of the partners building alongside AMD for our joint customers.”

About Cornelis

Cornelis delivers high-performance, scale-out and scale-up networking solutions that accelerate AI and HPC workloads. Cornelis technology enables lossless, congestion-free networking that reduces training time, improves inference, and maximizes compute utilization. From foundation model training to complex climate modeling and real-time analytics, Cornelis solutions power the most demanding workloads across commercial, academic, government, and cloud environments. With a focus on performance, scalability, and efficiency, Cornelis helps organizations achieve faster insights and greater return on infrastructure investments. Learn more at www.cornelis.com.