Cornelis Networks provides the CN5000 Omni‑Path platform, a 400 Gbps lossless, congestion‑free scale‑out interconnect for AI training and high‑performance computing workloads. Its credit‑based transport, adaptive routing and incast‑aware flow control eliminate packet loss and tail latency, delivering up to 2× higher message rates and near‑linear scaling for clusters of up to 500 K endpoints while remaining interoperable with all major GPUs, CPUs and accelerators.
Funding
$25M raised to dateRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
Founders
Product
Problem
AI training and high‑performance computing workloads are increasingly limited by network congestion and unpredictable tail latency, which cause GPU/CPU idle time, longer training cycles, and reduced overall system efficiency. Traditional Ethernet or InfiniBand fabrics react to congestion after it forms, leading to packet loss, head‑of‑line blocking, and scaling bottlenecks.
Solution
Cornelis Networks addresses these limitations with the CN5000 Omni‑Path platform, a lossless, congestion‑free scale‑out interconnect designed for AI and HPC at any size. The fabric uses fine‑grained adaptive routing, incast‑aware flow control, and credit‑based transport to prevent queue buildup and eliminate the need for reactive mechanisms such as PFC or ECN. Operating at 400 Gbps per link, the solution delivers sub‑microsecond tail latency, up to 800 M messages / sec, and up to 2× higher message rates compared with competing 400 Gbps Ethernet fabrics. By keeping all nodes continuously productive, CN5000 enables near‑linear training performance, faster inference, and higher application throughput across clusters ranging from a few racks to 500 K+ endpoints. The platform is built on open standards, interoperable with all major GPUs, CPUs, and accelerators, and includes a complete 400 G switch portfolio and optimized host‑side software.
Target Audience
Primary customers are large‑scale AI training clusters, high‑performance computing centers, and enterprise datacenters that require deterministic, low‑latency networking for compute‑intensive workloads such as deep‑learning, scientific simulation, and data‑intensive analytics.
Features
- Credit‑based lossless transport with dynamic adaptive routing for per‑packet path selection
- Incast‑aware flow control and link‑level replay to eliminate head‑of‑line blocking and packet drops
- 400 Gbps SuperNIC and full 400 G switch portfolio supporting up to 500 K+ endpoints
- Open‑source host and management software stack for easy integration and automation
- Universal interoperability (BYOP) with leading GPUs, CPUs, and accelerators
- Enhanced resiliency with error‑free delivery and no error‑driven retries
- Power‑efficient design that maximizes GPU power availability and reduces overall datacenter energy use