AI networking is making huge strides toward standardization but Ethernet isn’t ready to displace the leading incumbent InfiniBand yet. In this episode of the Tech Field Day Podcast, Tom Hollingsworth is joined by Scott Robohn and Ray Lucchesi to discuss the state of Ethernet today and how it is continuing to improve. The guests discuss topics such as the dominance of InfiniBand, why basic Ethernet isn’t suited to latency-sensitive workloads, and how the future will improve the technology.
Apple Podcasts | Spotify | Overcast | Amazon Music | YouTube Music | Audio
Ethernet is not Ready to Replace InfiniBand Yet
AI networking is making huge strides toward standardization but Ethernet isn’t ready to displace the leading incumbent InfiniBand yet. In this episode of the Tech Field Day Podcast, Tom Hollingsworth is joined by Scott Robohn and Ray Lucchesi to discuss the state of Ethernet today and how it is continuing to improve. The guests discuss topics such as the dominance of InfiniBand, why basic Ethernet isn’t suited to latency-sensitive workloads, and how the future will improve the technology.
InfiniBand has been the dominant technology for AI networking since NVIDIA asserted itself as the leader in the market. The reasons for this a varied. NVIDIA acquired the technology from their 2019 acquisition of Mellanox. InfiniBand has been used extensively in high performance computing (HPC) systems for a number of years. Using it for AI, which is a very hungry application, was a natural fit. Since then, InfiniBand has continued to be the preferred solution due to low latency and the lossless nature of packet exchange.
Companies such as Cisco, Broadcom, and Intel have championed the use of Ethernet as an alternative to InfiniBand for GPU-to-GPU communications. They’ve even founded a consortium dedicated to standardizing Ethernet fabrics focused on AI. However, even though Ethernet is a very flexible technology it’s not uniquely suited to AI networking in the same way that InfiniBand has shown. Lossy transmissions and high overhead are only two of the major issues that plague standard Ethernet when it comes to latency-sensitive information exchange. The Ultra Ethernet Consortium was founded to provide mechanisms to make Ethernet more competitive in the AI space but it still has a lot of work to do in order to standardize the technology.
The future of Ethernet is bright. InfiniBand is seemingly being put into maintenance mode as even NVIDIA has started to develop Ethernet options with Spectrum-X using Bluefield-3 DPUs. Cloud providers offering AI services are also mandating the use of standard cost-effective Ethernet over proprietary InfiniBand. AI workloads are also undergoing significant changes due to the nature of infrastructure catching up to their needs. As the technology and software continue to develop there is no doubt that Ethernet will eventually return to being the dominant communications technology. However, that change won’t happen for a few years to come.
Podcast Information:
Tom Hollingsworth is a Networking and Security Specialist at Gestalt IT and Event Lead for Tech Field Day. You can connect with Tom on LinkedIn and X/Twitter. Find out more on his blog or on the Tech Field Day website.
Ray Lucchesi is the president of Silverton Consulting and the host of Greybeards on Storage Podcast. You can connect with Ray on X/Twitter or on LinkedIn. Learn more about Ray on his website and listen to his podcast.
Scott Robohn is the VP of Technology at Cypress Consulting and Cofounder of Network Automation Forum. You can connect with Scott on X/Twitter or on LinkedIn. Learn more on the Network Automation Forum.
Thank you for listening to this episode of the Tech Field Day Podcast. If you enjoyed the discussion, please remember to subscribe on YouTube or your favorite podcast application so you don’t miss an episode and do give us a rating and a review. This podcast was brought to you by Tech Field Day, home of IT experts from across the enterprise, now part of The Futurum Group.
