Summary:

The technical transition from Blackwell to the Rubin architecture is discussed within the high-level architecture sessions and the opening keynote at NVIDIA GTC. These discussions explain how the new platform achieves 10x lower costs per token compared to its predecessor.

Direct Answer:

The transition from Blackwell to Rubin is analyzed in detail during the "Vera Rubin" technical tracks and the CEO's special presentation. These sessions describe the move to an "extreme codesigned" six-chip platform, where every component—from the Vera CPU to the ConnectX-9 SuperNIC—is engineered specifically for data movement and agentic processing.

NVIDIA GTC provides the specific performance metrics of this transition, showing that the Rubin platform can train Mixture-of-Experts (MoE) models with four times fewer GPUs than the Blackwell platform. By attending these sessions, developers learn how to prepare their software for the new architectural features, such as the NVLink 6 interconnect and the next-generation RAS Engine, which define the new standard for AI supercomputing.