At the Nvidia GPU Technology Conference (GTC) 2025, CEO Jensen Huang provided a glimpse into the future of AI computing with the announcement of Nvidia's next-generation high-end datacenter GPUs and CPUs, codenamed Vera, Rubin, and Rubin Ultra. These new architectures promise to significantly enhance performance and efficiency in data centers, marking a pivotal step forward in Nvidia's AI capabilities.Understanding Vera, Rubin, and Rubin UltraTo clarify the lineup, Vera is a CPU architecture, Rubin is a GPU architecture, and Rubin Ultra is a refreshed, more powerful version of Rubin. Each component is designed to work synergistically to deliver unparalleled performance.Vera: Named after the pioneering American astronomer Vera Rubin, known for her research on dark matter, Vera represents Nvidia's first Arm-compatible CPU architecture since the announcement of Grace in 2021. This CPU will feature 88 custom-designed Arm cores, leveraging Simultaneous Multithreading (SMT) to achieve a thread count of 176 per socket. Expected to arrive late next year, Vera will include integrated NVLink chip-to-chip connectivity for seamless interfacing with Nvidia's upcoming Rubin GPUs.Rubin: Drawing heavily from the Blackwell design architecture, Rubin GPUs will feature two reticle-limited dies capable of up to 50 petaFLOPS at FP4 precision, complemented by 288 GB of HBM4 memory providing 13 TB/s of bandwidth. Like Blackwell and Blackwell Ultra, Rubin will be packaged as a Superchip and deployed in Nvidia's rackscale NVL144 chassis. Notably, while the physical number of GPU dies remains consistent with previous models, Nvidia is now counting each die within the package as a separate GPU.Rubin Ultra: Set to debut in late 2027, Rubin Ultra will take performance to the next level by doubling the number of GPU dies and HBM modules to four and sixteen, respectively. Each Rubin Ultra package is projected to exceed 100 petaFLOPS of FP4 performance and incorporate 1 terabyte of faster HBM4e memory. A rack-scale system will house 144 of these packages along with an unspecified number of Vera CPUs, all within a 600 kW power consumption and thermal output rating. This configuration is expected to deliver 15 exaFLOPS of FP4 inference performance and 5 exaFLOPS of FP8 compute for training.Performance and Technological AdvancementsCompared to the GB300 NVL72, the Vera-Rubin NVL144 is projected to deliver a 3.3x increase in floating-point performance, topping 3.6 exaFLOPS of dense FP4 for inference and 1.2 exaFLOPS of FP8 compute for training. The system will also incorporate Nvidia's 6th-gen NVLink switch fabric, providing an aggregate 260 TB/s (1.8 TB/s per die) of interconnect bandwidth, and will utilize upcoming 1.6 Tbps ConnectX-9 NICs.Additional Announcements at GTCBeyond the Vera, Rubin, and Rubin Ultra architectures, Nvidia also unveiled:The Llama Nemotron family of reasoning models, designed for integration into agentic AI systems. The initial models, Nano and Super, are based on Meta's Llama 3.1 8B and 3.3 70B, pruned and fine-tuned for on-demand reasoning capabilities. These models are available as Nvidia Inference Microservices (NIMs) or via Hugging Face.An AI-Q Blueprint, an open-source framework for building complex agentic AI services capable of ingesting and processing information from multiple sources or databases.Challenges and Future ImplicationsThe immense power consumption of Rubin Ultra systems, potentially reaching 600 kW per rack, raises questions about the readiness of existing data center facilities to support such dense configurations. The graphic shared during Huang's keynote depicted a densely packed rack with systems slotting vertically into the cabinet, reminiscent of some HPC clusters from Lenovo and HPE Cray.To manage the increased data throughput, Nvidia plans to transition to faster NVLink7 interconnects for chip-to-chip communications while maintaining the 1.6 Tbps ConnectX-9 NIC for scale-out communications.Conclusion: A Bold Step into the Future of AINvidia's unveiling of the Vera, Rubin, and Rubin Ultra architectures at GTC 2025 signals a bold step into the future of AI computing. These advancements promise significant performance gains and enhanced capabilities for AI applications. However, the industry must also address the challenges associated with increased power consumption and infrastructure requirements to fully realize the potential of these groundbreaking technologies. As AI continues to evolve, Nvidia's innovations will undoubtedly play a crucial role in shaping the landscape of computing technology.
