Nvidia's Vera-Rubin Platform Obsoletes Current AI Iron 6 Months Ahead of Launch (www.nextplatform.com)

Nvidia has unveiled its upcoming Vera-Rubin platform, which promises significant advancements in AI computing capabilities, set to launch in the second half of 2026. This new rackscale system features the VR200 NVL72 configuration, boasting 72 GPU sockets linked by NVLink, alongside 36 CPU sockets. The platform offers a remarkable 10X reduction in inference costs per token for mixture of expert (MoE) models, alongside a 4X reduction in the number of GPUs needed for training when compared to its predecessor, the Blackwell GB200 system. This leap forward illustrates Nvidia's commitment to pushing the boundaries of AI hardware, enhancing performance while addressing industry demands for cost efficiency. Key technical innovations include the Rubin GPU, which combines eight stacks of HBM4 memory to deliver an unprecedented 22 TB/sec of memory bandwidth, significantly boosting token processing capabilities. Additionally, the new Vera CPU introduces spatial multithreading and improved cache architecture, enabling faster data sharing between CPUs and GPUs. As Nvidia gears up for production, the strong interest from major clients like AWS, Google Cloud, and Microsoft Azure highlights the competitive landscape, particularly as these companies explore their own accelerator development. The Vera-Rubin platform's advancements could enhance AI efficiency and performance at a pivotal moment in the rapidly evolving AI/ML landscape.