News
Nvidia reveals Vera CPU board architecture with dual-processor configuration at GTC 2026
Wednesday, March 18, 2026 at 11:46 PM
During the GTC 2026 event, Nvidia showcased the internal architecture of the Vera CPU board, revealing that each board is configured with two Vera CPUs. The demonstration included the removal of cold plates to provide a detailed view of the hardware layout designed for high-performance AI infrastructure.
Context
At GTC 2026, Nvidia officially detailed the Vera CPU board architecture, a cornerstone of its new Vera Rubin platform. The board features a dual-processor configuration where each unit carries two Vera CPUs, moving away from the single-chip Grace superchip design to a more modular, disaggregated architecture. Built on a 3nm process with 88 custom Olympus cores per chip, the Vera CPU provides 2X the performance of the previous generation and supports 1.5 TB of LPDDR5X memory. These systems are designed for liquid-cooled AI factories, with each rack integrating up to 256 Vera CPUs to handle the high-throughput demands of agentic AI and reinforcement learning.
During the event, Nvidia CEO Jensen Huang emphasized the strategic shift toward rack-scale computing, stating that "AI needs CPUs for tool use, and Vera CPU was designed just perfectly for that sweet spot." This transition is expected to turn the CPU business into a multi-billion dollar segment for the company as it projects total computing demand to exceed $1 trillion by 2027. The architecture’s reliance on NVLink-C2C interconnects allows for 1.8 TB/s of bidirectional bandwidth, ensuring the CPU can feed data to Rubin GPUs without bottlenecks, drastically reducing the cost per token for large-scale inference.
Sources (7)
NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer | NVIDIA Technical BlogNext Gen Data Center CPU | NVIDIA Vera CPUNVIDIA Vera Rubin Opens Agentic AI Frontier | NVIDIA NewsroomNVIDIA at GTC 2026: AI Expansion and Strategic Partnerships By Investing.comVera Rubin – Extreme Co-Design: An Evolution from Grace Blackwell OberonNvidia updates data center roadmap with Rosa CPU and stacked Feynman GPUs — optical NVLink, Groq LPUs with NVFP4, and NVLink also on deck | Tom's HardwareNVIDIA GTC 2026: Rubin GPUs, Groq LPUs, Vera CPUs, and What NVIDIA Is Building for Trillion-Parameter Inference - StorageReview.com
Related Companies
Nvidia
NVDA