DriveNets, WhiteFiber and WEKA successfully connected GPU clusters at two geographically separated data centers to operate as one AI supercluster. The video explores how this scale-across approach helps overcome power and capacity constraints, alongside network planning considerations and plans to expand to more sites.
𝗞𝗲𝘆 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆𝘀
- Plan the network ahead: use shallow buffer for intra-site connectivity and deep buffer for cross-site connectivity.
- Provide sufficient capacity, preferably dark fiber, between the sites, and use a scheduler that spreads the workload across different systems.
- • When a data center runs out of power, add GPUs in another data center up to 50 and even 100 miles away. Both GPU clusters then work as a single cluster.
- For latency-sensitive inference, place GPUs closer to customers at the edge, and measure time to first token and time between tokens, not just round-trip delay.
- WhiteFiber runs 2 sites today and plans 3 or more in a mesh or hub-and-spoke model, combining smaller modular capacity into one larger fabric.
𝗥𝗲𝗮𝗱 𝘁𝗵𝗲 𝗳𝘂𝗹𝗹 𝘀𝘁𝗼𝗿𝘆
https://www.whitefiber.com/blog/cross-data-center-networking-solution
---------------------------------------------------------------------
𝗟𝗲𝗮𝗿𝗻 𝗺𝗼𝗿𝗲 𝗮𝗯𝗼𝘂𝘁 𝗗𝗿𝗶𝘃𝗲𝗡𝗲𝘁𝘀
https://drivenets.com/
DriveNets builds high-performance networking solutions for large-scale AI infrastructure and service-provider networks. DriveNets’ Network Cloud powers production networks at global operators such as AT&T and Comcast, while its AI Fabric solutions optimize the performance and time to first token (TTFT) in multi-vendor AI environments at foundation model labs, NeoClouds, and enterprises. DriveNets combines high-performance networking, end-to-end automation, and full-stack AI cluster optimization to remove network bottlenecks, maximize GPU utilization, accelerate AI cluster deployments, and improve token economics.
𝗪𝗵𝗲𝗿𝗲 𝘁𝗼 𝗳𝗶𝗻𝗱 𝗗𝗿𝗶𝘃𝗲𝗡𝗲𝘁𝘀
🔗 LinkedIn: https://www.linkedin.com/company/drivenets
🔗 X: https://x.com/drivenets
D'autres épisodes de "CloudNets"



Ne ratez aucun épisode de “CloudNets” et abonnez-vous gratuitement à ce podcast dans l'application GetPodcast.








