NVIDIA Outlines Vera Rubin Ultra NVL576 Roadmap for a 576-GPU Compute Domain
Редакция ЦОД:Главное
NVIDIA used OCP 2026 on October 22 to outline its planned Vera Rubin Ultra NVL576 architecture, which is intended to link 576 GPUs into a single compute domain using co-packaged optics (CPO) and NVLink 7. The platform remains a roadmap proposal, with performance and memory specifications presented as targets rather than confirmed commercial product characteristics.
- NVIDIA plans to use CPO and NVLink 7 to link 576 Rubin Ultra GPUs into a single compute domain.
- The proposed design extends high-speed GPU interconnects across multiple racks rather than confining the domain to one rack.
- The Vera Rubin Ultra Compute Tray is planned with two Rubin Ultra GPUs, one Vera CPU and fully liquid cooling.
- NVIDIA's published performance, memory and bandwidth figures are roadmap targets, not confirmed commercial specifications.
- Potential infrastructure implications include greater demand for liquid cooling, high-density racks and rack-to-rack optical interconnects, if the platform is commercialized.
NVIDIA has disclosed further plans for Vera Rubin Ultra NVL576, an AI platform architecture intended to aggregate 576 GPUs into one compute domain through CPO and NVLink 7. Presented at OCP 2026 on October 22, the design would extend GPU connectivity beyond a single rack through optical and NVLink interconnects.
The company showed an NVLink 7 CPO Switch Tray and a Vera Rubin Ultra Compute Tray. Each Compute Tray is designed to integrate two Rubin Ultra GPUs and one Vera CPU, with fully liquid cooling. The stated direction is a multi-rack architecture for AI workloads rather than a system bounded by the scale of one server or rack.
NVIDIA positions the NVL576 architecture for AI models exceeding 10 trillion parameters. Its roadmap targets include 15 exaflops of FP4 inference performance and 5 exaflops of FP8 training performance. The company also cited up to 1 TB of HBM4E per Rubin Ultra GPU, up to 365 TB of total HBM across the platform, and up to 4.6 PB/s of aggregate memory bandwidth. A comparison with Blackwell GB300 NVL72 indicates approximately 14 times higher performance, though the supporting methodology is not provided in the available material.
For data-center operators and AI-infrastructure suppliers, the relevant shift is architectural: large GPU domains would depend on optical connectivity between racks alongside direct-liquid-cooled compute hardware. If these designs are commercialized, they could support future demand for scalable liquid-cooling systems, high-density rack infrastructure, optical switching and modular components designed for serviceability.
The announcement does not establish deployed capacity, committed procurement or a commercial launch timetable. It should therefore be read as a roadmap signal rather than evidence of operational AI-campus infrastructure demand.
The proposed architecture points to AI infrastructure scaling beyond a single server or rack toward optically interconnected, multi-rack compute domains. If commercialized, it could increase requirements for scalable liquid cooling, high-density rack infrastructure, rack-to-rack optical interconnects and standardized modular service components.
- TechNews 科技新報 / КитайскийПервоисточник →
Материал подготовлен редакцией ЦОД: Главное на основе указанных первоисточников.