Edgecore Networks has deployed open-networking data center switches at HyperVerge to support AI training and IP storage workloads, using 100G/400G Ethernet and SONiC to target high-bandwidth, low-latency storage traffic inside the cluster. The setup is built on Broadcom switching hardware and is intended to keep latency tightly controlled while allowing HyperVerge to self-configure the environment quickly.
Edgecore describes the design as a datacenter-based 100G/400G solution running on open networking for “hi-bisectional” data-storage traffic. In HyperVerge’s environment, the switching fabric is a 100G RoCEv2 lossless RDMA network running SONiC, used to connect a GPU cluster to shared NVMe storage that HyperVerge has unified into a single clustered parallel file system.
For data center operators, the interesting part isn’t the port speed headline—it’s the specific choice of RoCEv2 with a lossless fabric for east-west storage traffic. If HyperVerge is pushing AI training data over the network instead of relying on node-local storage, then jitter and congestion control become first-order problems: a storage bottleneck doesn’t just slow jobs down, it can leave GPUs underutilized.
“For AI training workloads, storage bandwidth and latency are as critical as compute. Any bottleneck directly translates into underutilized GPUs,” said Manideep Kolla, Head of Identity AI at HyperVerge. “At HyperVerge, we unified all our NVMe storage into a single clustered parallel file system, and Edgecore’s 100G RoCEv2 lossless RDMA fabric powered by SONiC enables us to deliver this storage at near-local NVMe latency and bandwidth across our GPU cluster over the network.”
HyperVerge is based in Bengaluru and builds identity verification tools including Video KYC, OCR-based ID data extraction, and Face Authentication aimed at deepfake prevention. The company says its platform reaches 99.5% accuracy and can operate on 2G connections, and it uses AI models spanning face recognition, OCR, and custom LLMs for identity verification, digital KYC, and underwriting intelligence. HyperVerge says it processes hundreds of millions of onboarding journeys annually.
“By using Edgecore’s high-throughput data center design coupled with the SONiC open-networking NOS, the team was able to quickly design a solution for the HyperVerge datacenter storage cluster,” said Nanda Ravindran, VP of Product Management and Planning at Edgecore Networks.
Edgecore Networks and HyperVerge did not disclose switch model numbers, topology details, or deployment scale.
Source: Edgecore Networks
