𝗜𝗻𝘁𝗲𝗹 𝗢𝗺𝗻𝗶-𝗣𝗮𝘁𝗵 𝗥𝗲𝘁𝘂𝗿𝗻𝘀 𝘁𝗼 𝗖𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲 𝗡𝘃𝗶𝗱𝗶𝗮
Cornelis Networks is bringing Intel Omni-Path back to life. This technology will now connect U.S. Department of Energy supercomputers. It serves as a direct alternative to Nvidia InfiniBand.
Intel stopped making Omni-Path in 2019. Now, Cornelis Networks uses it to provide 400Gbps speeds. This move helps the Department of Energy avoid relying only on Nvidia hardware.
Key details of the new hardware:
• The Cornelis CN5000 switch supports 200 ports. • Each port runs at 400Gbps. • Latency stays under 100 nanoseconds per hop. • The tech uses a direct memory access model. • This model avoids TCP/IP overhead to keep speeds high.
Why this matters for AI and Supercomputing:
Large scale AI training needs fast connections between chips. InfiniBand is the current leader. However, Omni-Path claims lower tail latency. Low latency is vital for model training tasks.
The Department of Energy plans to use this for exascale systems between 2027 and 2028.
Challenges ahead:
Nvidia has a massive advantage with its software ecosystem. Omni-Path requires specific drivers and libraries. Success depends on how well the software works with existing AI tools.
Cornelis Networks also faces competition from the Ultra Ethernet Consortium. This group includes AMD and Intel. The race to dominate supercomputer networking is heating up.
Optional learning community: https://t.me/GyaanSetuAi