Cornelis CN5000
The world's first lossless, congestion-free scale-out network
What is CN5000?
CN5000 is a 400 Gb/s end-to-end Omni-Path fabric built to maximize AI and HPC performance. By integrating SuperNICs, Switches, Director Class Switches, software, and cabling into a single architecture, CN5000 delivers lossless operation, adaptive routing, intelligent congestion control, and deep telemetry to keep data flowing and accelerators productive. Cornelis doesn't assemble networking solutions from multiple sources–it controls the entire fabric stack. From silicon to software, every component is designed and tuned to work together, enabling higher application performance, greater operational efficiency, and improved utilization of valuable compute resources.
- 2XHigher message rates*
- +30%Latency*
- 30%Faster simulation times+Cited HPC comparisons
*Compared with NDR InfiniBand, based on an internal Cornelis benchmarking study. Contact Cornelis for more information.
Keep your network moving at scale
Spend less time
waiting
Higher message rates and lower latency help move data faster between processors and accelerators. AI workloads and HPC simulations spend less time waiting on communication, helping teams reach results sooner.
Keep applications moving
as you scale
A 400G fabric designed for up to 500,000 endpoints gives large AI and HPC systems room to grow. Sustained throughput and controlled tail latency under load help applications maintain predictable performance as traffic increases.
Fit the fabric
to your environment
One family with SuperNICs, Edge and Director Class switches, software, cabling, and air or liquid cooling options support different data center designs and give more deployment flexibility.
Designed as one system,
from silicon to software
CN5000 ships as a complete fabric. Cornelis builds a hardware and software system, silicon, board, switch and stack, so every layer is tuned against the same architecture. By tuning every layer against one architecture, Cornelis brought CN5000 to market on first-pass A0 silicon with measured throughput close to line rate.
SuperNIC
PDFThe CN5000 SuperNIC delivers one or two 400 Gb/s ports on a low-profile PCIe Gen 5 x16 adapter. Credit-based flow control, enhanced congestion control, and high injection rates minimize network stalls and congestion, helping AI and HPC workloads sustain peak performance.
Learn more about the SuperNIC
Switch
PDFThe CN5000 Switch delivers 48 ports of 400 Gb/s connectivity and 38.4 Tb/s of full-duplex switching bandwidth in a high-density platform. Up to 32 ports can be split into dual 200 Gb/s links for flexible deployment. Available in air- and liquid-cooled configurations, it combines enhanced congestion control with fine-grained adaptive routing to maintain performance and efficiency under demanding AI and HPC workloads.
Learn more about the Switch
Director Class Switch
PDFThe CN5000 Director Class Switch (DCS) scales to 576 ports at 400 Gb/s and up to 230.4 Tb/s of aggregate switching bandwidth in a single chassis. Ports are enabled in 48-port increments, providing flexible growth without redesigning the network. Its integrated two-tier fat-tree architecture eliminates large numbers of optical cables and external switches, reducing deployment complexity, simplifying operations, and lowering effective cost per port.
Learn more about the Director Class Switch
Copper and Optical Cables
Cornelis offers low cost, high reliability, and low power consumption cables for purpose-built high-performance networking.
Learn more about Copper and Optical Cables
OPX Software
Cornelis OPX is an open software stack that spans host drivers, fabric services, and switch management. Built on the OFI libfabric framework, it provides broad ecosystem compatibility and a standards-based foundation for AI and HPC. MPI, OpenSHMEM, and RCCL applications run unchanged, enabling seamless adoption and portability across environments.
Learn more about OPXCN5000 compared with InfiniBand and Ethernet
| Attribute |
|
|
|
|---|---|---|---|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Systems, platform and channel partners
See the full Cornelis partner ecosystem
Distribution and Federal Integrators
Pricing and availability
CN5000 is delivered through server OEMs and channel partners.
Pricing depends on the system configuration, not a single list price.
How CN5000 is sold
Server OEMs and channel partners deliver CN5000 as part of configured systems. Federal buyers can work through approved integrators.
What shapes a quote
Node count and endpoint scale
Switch type and scale: Edge or Director Class
Cabling: passive copper, active copper, or active optical
Cooling: air, hybrid, or liquid
Availability
Cornelis-designed silicon gives CN5000 a direct path from architecture to production. Eligible configurations can ship within six weeks of order confirmation, helping teams move from planning to deployment sooner.*
*Lead times vary by configuration, quantity, inventory, and component availability. Contact Cornelis or your channel partner to confirm current availability.
- PORT BANDWIDTH400 Gb/sOne or two fabric ports per SuperNIC
- MEASURED THROUGHPUT385 Gb/sAverage across tested node pairs, 378 to 391, roughly 96 percent of line rate
- MEASURED LATENCY1.1 μsApproximate, node to node.
Measured results depend on system configuration, workload, topology, and test conditions.
Open software and interoperability
- An open source OPX software suite, including a libfabric provider based on the Open Fabrics Interfaces (OFI) framework.
- Applications using Open MPI, MPICH, MVAPICH2, NCCL, and RCCL run without modification.
- OpenBMC-based chassis management and Redfish support, with in-band and out-of-band management options.
- Supported Linux distributions: Red Hat Enterprise Linux, Rocky Linux, and Ubuntu.
Frequently asked questions
Everything about the CN5000 fabric in one place — architecture, positioning against InfiniBand and Ethernet, roadmap and procurement.
-
What is the Cornelis CN5000?
CN5000 is a 400 Gb/s end-to-end Omni-Path® fabric for AI and HPC clusters, and an alternative to InfiniBand orEthernet for connecting GPU and CPU nodes at scale. It includes SuperNICs, 48-port Switches, DirectorClass switches of up to 576 ports, copper and optical cabling, and the open-source OPX software suite. It is theshipping generation of what Cornelis calls the active compute fabric: a fabric built to compute on data in flight,not only to move it between endpoints. -
How does CN5000 differ from InfiniBand or Ethernet?
The difference is architectural rather than a matter of speed. InfiniBand applies credit-based flow control at the link level while fabric-level congestion persists at scale. Ethernet is lossy by design and relies on pause frames and explicit congestion notification, which introduce their own penalties. CN5000 is lossless and congestion-freeat the fabric level, and routes adaptively at fine grain in real time rather than hashing flows to fixed paths. -
What is the difference between CN5000, CN6000 and CN7000?
CN5000 is the shipping 400G generation and delivers active transport: lossless delivery, adaptive routing and congestion elimination. CN6000 is the next step, an 800G multi-protocol SuperNIC that adds RoCEv2 and Ultra Ethernet compatibility while the end-to-end fabric remains Omni-Path. CN7000 extends the architecture into the GPU pod and brings Distributed Compute Fabric capabilities, programmable compute at every port, into the design. -
What interconnect products does Cornelis Networks make, and how are they priced?
Cornelis builds three families. CN5000 is the shipping 400G end-to-end Omni-Path fabric of SuperNICs, Edge and Director Class switches, cabling and OPX software. CN6000 is an 800G multi-protocol SuperNIC adding RoCEv2 and Ultra Ethernet. Omni-Path 100 remains available for existing 100G estates. Cornelis sells through OEM and channel partners rather than direct, so pricing is quoted per configuration against node count, switch topology, cabling and cooling. -
Is CN5000 a good fit for a new AI training cluster with thousands of GPUs?
It is built for that traffic. Training is dominated by collective operations, where every node waits for the slowest message, so tail behaviour matters more than peak bandwidth. CN5000 is lossless and congestion-free with incast-aware flow control and fine-grained adaptive routing. It interoperates with AMD, Intel and NVIDIA accelerator platforms, and the modular 576-port Director Class Switch adds ports without adding a switching tier. -
How does CN5000 hold performance under sustained load?
Congestion, incast and shifting traffic patterns are what erode throughput and stretch tail latency in production. CN5000 answers with credit-based and incast-aware flow control, dynamic fine-grained adaptive routing, static dispersive routing, fine-grained Quality of Service, optimized link protection and advanced error correction, with telemetry to show where load actually sits. TACC verified the result through formal acceptance testing rather than a benchmark run. -
Is Cornelis a long-term vendor for national laboratory procurement?
The evidence is public. CN5000 is in production on Lynx at Lawrence Livermore for the NNSA and listed on the Top500, and TACC completed National Science Foundation approved acceptance on Stampede3. Cornelis reaches federal buyers through integrators including CTG Federal and TVAR Solutions, ships on OEM platforms from Lenovo, Dell, Supermicro and xFusion, contributes to libfabric adopted by the Ultra Ethernet Consortium, participates in the UALink Consortium, and has published a roadmap through CN6000 and CN7000.
Ready to evaluate CN5000?
Talk to an engineer about your cluster, your workloads and your scaling targets.
Scales fluidly to 500k+ endpoints