Nvidia is looking to meet demand from network engineers for diversity at the transport layer by extending support for Google’s hardware transport protocol, Falcon.
Announced at the OCP Global Summit in San Jose, the move will see future BlueField data processing units and ConnectX-10 SuperNICs support the transport, allowing Nvidia and Google hardware to communicate over a common transport while applications retain standard remote direct memory access (RDMA) programming interfaces.
Falcon was unveiled at Global Summit back in 2023 and is designed to provide reliable, high-performance, and low-latency transport for AI infrastructure. It’s one of a growing number of options at the transport layer, with Nvidia already supporting the OpenAI-developed Multipath Reliable Connection (MRC).
Google’s Falcon option is now supported on Nvidia Spectrum-X Ethernet offerings, with Nvidia touting it as a means to help distribute traffic and aid with workloads to “make better use of available bandwidth and maintain predictable performance as Ethernet fabrics scale.”
Joe DeLaere, Nvidia’s product lead for AI infrastructure solutions, said during a press briefing that network engineers working on mammoth AI clusters can use Falcon to “benefit from multipath load balancing, which spreads traffic to reduce bottlenecks, programmable congestion control that adjusts traffic to limit network delays, hardware loss recovery that keeps data moving when packets drop, and per-connection security that protects data while in transit.”
“All of these help customers scale AI and storage with more predictable performance and secure, reliable data delivery,” the exec added.
The rise in protocol options comes as a result of traditional transports like TCP and RoCEv2 (RDMA over Converged Ethernet) hitting performance walls when it comes to connecting tens or hundreds of thousands of GPUs. Hyperscalers and Nvidia customers want lossless configurations to avoid packet drops while ensuring optimal routing efficiency at higher AI-induced speeds.
In the years since Falcon’s unveiling, transport alternatives include the Ultra Ethernet Consortium-developed Ultra Ethernet Transport (UET) and BXI-V3 from supercomputing giant Bull. Meanwhile, engineers from Nvidia rival AMD recently lent their expertise to help Meta develop its own transport protocol, the aptly named MetaRoCE.
Nvidia confirmed that beyond Falcon, support would be extended to more custom transport options in the future.
“Spectrum-X is completely transport agnostic, and you've probably seen previous announcements that we had earlier this year with MRC support, so it's very likely that you'll see other transport support coming in the near future,” Nvidia senior director of marketing for networking platforms Brian Sparks said in a response to an SDxCentral query.