An Active Compute Fabric™ is an AI and HPC interconnect that works for the workload rather than only carrying it. Where InfiniBand and Ethernet forward packets between endpoints, it computes at every hop, so congestion control and collectives run in the network instead of on your accelerators. One open architecture, across scale-up and scale-out.
The shift: from passive transport to active compute
Compute, storage, and memory all evolved to understand the workload. The network is now making the same shift, from moving data to acting on it. That shift is architectural: it makes programmable compute in the silicon, at every hop.
An active compute fabric is a purpose-built interconnect architecture that makes the network work for AI and HPC workloads, not just carry data. An active compute fabric brings three cumulative levels of capability to the fabric, with each level building on the one before it, from adapting data movement, to accelerating communication, to adding programmable compute in the fabric.
Adapts how to move data. Credit-based flow control and dynamic, fine-grained adaptive routing helps manage congestion in the fabric and respond to changing traffic conditions.
Active acceleration
Brings communication work into the fabric. Adds native collective acceleration in the fabric, reducing communication work in host software.
Active compute
Adds programmable compute at every port through Distributed Compute Fabric™ (DCF) capabilities. Executes workload-specific operations in transit, extending the network into a programmable compute layer.
High-performance infrastructure
Active Compute Fabric closes the gap between hardware capability and actual results, and then opens an optimization layer that passive fabrics cannot provide.
>2x*higher message rates than comparable 400 Gb/s solutions.
+30%*lower latency than InfiniBand.
30%better performance per networking dollar, maximizing ROI.
*Compared with NDR InfiniBand, based on an internal Cornelis benchmarking study. Contact Cornelis for more information.
One architecture across scale-up and scale-out
Active compute fabric unifies scale-up and scale-out under one open architecture.
Scale-up binds accelerators into one shared-memory domain. Scale-out links tens of thousands of nodes into one low-latency fabric. Most organizations run a different architecture at each layer, with two management planes, and every pod-to-pod interaction crosses the seam between them. CN5000 establishes the scale-out foundation today on Omni-Path. CN6000 extends that path with RoCEv2 and Ultra Ethernet compliance through a multi-protocol SuperNIC, while the end-to-end fabric remains Omni-Path. CN7000 will extend the architecture into scale-up through native UALink and ESUN support.
Key capabilities
Programmable compute at every port hosts a composable set of acceleration functions on one silicon investment. CN7000 is designed to execute them at 3.2 TB/s with 96 programmable cores per switch and 8 per adapter.
Forward-looking products are based on design targets and subject to change.
A four-stage in-fabric pipeline for prefill-to-decode KV cache transfer in disaggregated serving. For agentic and long-context workloads, it is designed to reduce host-side work and help shorten time to first token.
MoE Expert Dispatch
Content-addressed deduplication and load-aware routing for all-to-all token traffic. Designed to reduce duplicate transfers and communication overhead for MoE inference and training.
In-Network Collectives
Ultra Ethernet Consortium (UEC)-conformant AllReduce, AllGather, and ReduceScatter with in-network quantization. Designed to move collective work into the fabric and reduce communication overhead for AI training and other collective-bound workloads.
MPI/HPC Offload
NIC-resident schedule replay and hardware completion for persistent collectives. Designed to reduce recurring host-side communication work in MPI and HPC workloads.
PGAS/SHMEM
Hardware completion and atomic combining for NVSHMEM and OpenSHMEM at high message rates. Designed to support one-sided communication for MoE and HPC workloads.
Checkpoint/Goodput
In-fabric checkpoint, restore and failure recovery for large jobs. Wins back the calendar time lost to failures at scale.
Active compared with passive
Designed for a single solution that offers scale-up and scale-out: Active Compute Fabric. The distinction is architectural, not company-specific.
Purpose-built for the converged AI and HPC workload class
02
Congestion handling
Credit-based at the link level, fabric-level congestion persists
Inherently lossy, PFC and ECN carry their own penalties
Lossless and congestion-free at the fabric level
03
Adaptive routing
Static or limited adaptive
ECMP, hash-based rather than dynamic
Fine-grained adaptive routing, recalculated in real time
04
Collective communication
Handled by host software
Handled by host software
Faster native collective acceleration compared with RoCE
05
In-network compute
Fixed reduction operators only, no per-port programmability
None shipping. No programmable substrate exists in shipping Ethernet Switch silicon
Programmable compute in every Switch and SuperNIC, defined in firmware
06
Scale-up and scale-out
Separate proprietary interconnect at scale-up, two management planes
Fragmented, different vendors and architectures at each layer
One architecture across both layers, open to any accelerator
07
Lock-in
Single vendor controls the stack
Multi-vendor, fragmented standards
Open protocols, open-source software, standards-based hardware
Frequently asked questions
How is an Active Compute Fabric different from a DPU or SmartNIC?
A DPU or SmartNIC adds processing at an endpoint, often to offload work from the host. An Active Compute Fabric extends processing into the network itself. It adapts data movement and manages congestion in the fabric, then builds toward programmable compute at every port for workload-specific operations in transit.
Do applications need to be rewritten?
In most cases, application rewrites are not required. Cornelis hardware is designed to operate beneath supported application interfaces, communication libraries, and protocols, so applications can remain independent of the underlying fabric as infrastructure evolves.
Moving to a Cornelis fabric may still require updated drivers, providers, libraries, or configuration. Teams should validate supported versions and workload behavior for the selected product and software stack.
For CN5000, the OPX provider integrates with the Open Fabrics Interfaces framework and libfabric. Applications using supported versions of Open MPI, MPICH, MVAPICH2, NCCL, and RCCL can run without application rewrites, subject to supported configurations.
Which products deliver the Active Compute Fabric today?
Active Compute Fabric arrives in product steps. CN5000 ships today and delivers Active Transport as a scale-out fabric of SuperNICs, Switches, and Director Class Switches. CN6000 is arriving next with Active Acceleration and an 800 Gb/s multi-protocol SuperNIC. CN7000 is designed to deliver Active Compute and programmable compute at every port across scale-up and scale-out.
What is the Distributed Compute Fabric?
The Distributed Compute Fabric (DCF) is the in-network compute capability portfolio designed for CN7000. It includes In-Network Collectives, KV Cache Acceleration, MoE Expert Dispatch, MPI/HPC Offload, PGAS/SHMEM, and Checkpoint/Goodput. Designed on CN7000 for scale-up and scale-out.
Does an Active Compute Fabric work with Ethernet?
Yes. CN6000 is designed to bring the active-fabric approach into Ethernet and RoCEv2 environments through a multi-protocol SuperNIC. The end-to-end fabric remains Omni-Path, with RoCEv2 and Ultra Ethernet compatibility at the port level. This gives Ethernet-based deployments an entry point rather than requiring a rip-and-replace migration.
An active fabric works on the data, not just moves it
Full performance from the hardware you have, ready for the software you are yet to build.
