Cornelis has introduced a new reference architecture that pairs its CN6000 SuperNIC with AMD 6th Gen EPYC processors and AMD Instinct MI400 series GPUs, targeting disaggregated AI inference, large-scale training, and HPC simulation. The goal is to give infrastructure operators a blueprint for building an integrated compute-and-networking platform tuned for high-performance AI and HPC clusters.
The design centers on the CN6000 SuperNIC, which supports Omni-Path, RoCEv2, and Ultra Ethernet transport on a single adapter. Cornelis says that’s intended to let operators standardize on one NIC platform while retaining protocol flexibility. The adapter is built around a native PCIe 6.0 x16 host interface and is engineered for 800 Gbps of bandwidth. Cornelis also states that the CN6000 architecture is targeting more than 1.6 billion bidirectional messages per second.
Cornelis ties the reference architecture to upcoming AMD platforms, including AMD’s EPYC Venice processor, which AMD describes as being built on a 2-nanometer process with up to 256 cores per socket, and the AMD Instinct MI400 series GPUs for large-scale training and inference deployments. Cornelis’ angle is straightforward: as GPU counts scale and AI pipelines split across stages (including prefill, decode, and orchestration), network behavior becomes a direct limiter on time to first token, token throughput, and overall cluster utilization.
On performance, Cornelis points to pre-production simulations in AI training. In that modeling, Cornelis says a CN6000-based network completed AllReduce collectives roughly 24% faster than “standard Ethernet,” and reduced training time for a 250-billion-parameter model on a simulated 10,000-GPU cluster by about 13%. Those are meaningful deltas if they hold up in real deployments, but they’re still simulation results, and operators will want to see workload details, topologies, and tuning assumptions before treating them as capacity-planning inputs.
The broader CN6000 family includes the 800G multi-protocol SuperNIC, a 48-port Omni-Path switch, and a libfabric-based software stack that Cornelis says is compatible with AMD ROCm. Cornelis expects the CN6000 to be generally available in Q4CY26.
Cornelis also framed the reference architecture as an extension of its collaboration with AMD, which it said spans more than five years, more than 25 validated platform configurations, and nine OEM partners. The companies also operate a joint HPC Center of Excellence in Munich focused on validating performance on real workloads ahead of customer deployment.
“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,” said Lisa Spelman, CEO of Cornelis.
Source: Cornelis
