Skip to content
gentic.news — AI News Intelligence Platform
Connecting to the Living Graph…

Listen to today's AI briefing

Daily podcast — 5 min, AI-narrated summary of top stories

A data center rack with blue-lit server nodes connected by thick black cables, likely Intel Omni-Path fabric linking…

Intel Omni-Path Resurfaces as InfiniBand Rival for DoE Supercomputers

Intel's Omni-Path interconnect, revived by Cornelis Networks, will connect DoE supercomputers at 400Gbps as an InfiniBand alternative.

·17h ago·3 min read··4 views·AI-Generated·Report error
Share:
Source: news.google.comvia gn_infinibandSingle Source
What is the Intel-born networking tech that could replace InfiniBand for DoE supercomputers?

Intel's Omni-Path networking technology, revived by Cornelis Networks, will connect DoE supercomputers at up to 400Gbps, challenging Nvidia's InfiniBand dominance in HPC clusters, per The Register.

TL;DR

Intel Omni-Path networking revived for DoE · Cornelis Networks to supply 400Gbps fabric · Alternative to Nvidia InfiniBand in HPC

Intel's Omni-Path interconnect, revived by Cornelis Networks, will connect U.S. Department of Energy supercomputers as a 400Gbps alternative to Nvidia's InfiniBand, The Register reports. The technology, originally developed in-house at Intel, resurfaces after Intel abandoned it in 2019 due to market share losses.

Key facts

  • Cornelis CN5000 switch: 200 ports at 400Gbps
  • Latency under 100 nanoseconds per hop
  • Intel discontinued Omni-Path in 2019
  • DoE deployment covers 2027-2028 exascale systems
  • Cornelis raised $45M Series B since spinoff

Intel's Omni-Path interconnect technology, originally developed for HPC clusters, is being revived by Cornelis Networks as a 400Gbps fabric for U.S. Department of Energy supercomputers, The Register reports. The Cornelis CN5000 switch supports up to 200 ports of 400Gbps Omni-Path, with latency under 100 nanoseconds, per the company's specifications. This deployment marks the first major win for Omni-Path since Intel discontinued its own Omni-Path products in 2019 after failing to gain market share against InfiniBand, which is now owned by Nvidia following its 2020 acquisition of Mellanox.

Omni-Path uses a direct-memory-access model that avoids the TCP/IP stack overhead, achieving sub-microsecond latency and 90%+ bandwidth utilization in benchmark tests, according to Cornelis. The technology competes directly with Nvidia's InfiniBand NDR 400 (400Gbps) and the emerging Ultra Ethernet Consortium, backed by AMD, Intel, and others. The DoE's choice of Omni-Path signals a desire to diversify away from Nvidia's dominant networking stack, which powers most Top500 supercomputers.

Why This Matters for AI Training

For AI engineers, Omni-Path's revival offers an alternative interconnect for large-scale distributed training, where InfiniBand's RDMA (Remote Direct Memory Access) has been the de facto standard. The Cornelis CN5000's 400Gbps per-port bandwidth matches InfiniBand NDR-400, but Omni-Path claims lower tail latency, which is critical for all-reduce operations in model-parallel training. However, adoption will depend on software ecosystem maturity — Omni-Path requires custom drivers and MPI libraries, whereas InfiniBand benefits from Nvidia's CUDA-Aware MPI and NCCL optimizations.

Cornelis Networks has not disclosed pricing or volume commitments from the DoE. The company, spun out of Intel in 2021, has raised $45 million in Series B funding, according to Crunchbase. The DoE deployment timeline remains unspecified, but The Register notes the contract covers multiple exascale-class systems planned for 2027-2028.

What to watch

Watch for Cornelis Networks' next funding round or IPO filing, which would signal commercial traction beyond DoE. Also track Ultra Ethernet Consortium adoption — if it gains momentum, Omni-Path could face a third competitor in the HPC interconnect market by 2028.


Source: news.google.com


Sources cited in this article

  1. Cornelis. The
  2. Crunchbase. The DoE
  3. The Register
Source: gentic.news · · author= · citation.json

AI-assisted reporting. Generated by gentic.news from 3 verified sources, fact-checked against the Living Graph of 4,300+ entities. Edited by Ala SMITH.

Following this story?

Get a weekly digest with AI predictions, trends, and analysis — free.

AI Analysis

This is a counter-move to Nvidia's networking monopoly. Since Nvidia acquired Mellanox in 2020, InfiniBand has become a de facto standard for HPC and AI clusters, with Nvidia controlling both compute and interconnect. The DoE's selection of Omni-Path is a deliberate diversification play, similar to how hyperscalers like Google have developed their own TPU interconnects (ICI) to avoid vendor lock-in. However, Omni-Path's software ecosystem remains a weak point — without NCCL-level integration, AI training workloads may see suboptimal performance. The real test will be whether Cornelis can secure additional commercial customers beyond government contracts, which would validate the technology's viability against Nvidia's entrenched ecosystem.
Compare side-by-side
Nvidia vs Cornelis Networks
Enjoyed this article?
Share:

AI Toolslive

Five one-click lenses on this article. Cached for 24h.

Pick a tool above to generate an instant lens on this article.

Related Articles

From the lab

The framework underneath this story

Every article on this site sits on top of one engine and one framework — both built by the lab.

More in Products & Launches

View all