Google Unveils Virgo Scale-Out Fabric for AI Data Centers

Google has launched Virgo, a dedicated scale-out network fabric designed to connect massive clusters of GPUs and TPUs, capable of supporting 134,000 chips with 47 petabits per second of non-blocking bisection bandwidth [1]. The architecture marks a strategic shift for the company, separating high-speed accelerator-to-accelerator traffic from general-purpose data center networking to improve performance and fault tolerance in large-scale AI training and inference [1].

Architectural Shift for AI Clusters

Virgo functions as an "east-west" fabric, operating independently of Google’s existing Jupiter front-end network and the internal scale-up interconnects used within individual accelerator systems [1]. By utilizing high-radix switches—which house a high number of ports—Google has collapsed the network into a flat, two-layer topology [1]. This design replaces traditional three-layer networks, a change Google claims provides up to four times the bandwidth per accelerator and up to 40% lower unloaded fabric latency compared to previous configurations [2].

The fabric is engineered to handle the "inevitable" hardware failures inherent in clusters containing hundreds of thousands of chips [1]. Virgo employs a multi-plane design where each accelerator connects across multiple independent network planes; if one plane fails, the system retains 87.5% of its aggregate bandwidth rather than suffering a total connection loss [1]. To manage congestion, the architecture incorporates high-resolution telemetry and CSIG in-band signaling, allowing the network to detect microbursts and steer traffic around bottlenecks in real-time [1].

Scaling Beyond the Data Center

Google is positioning Virgo as the foundation for its next generation of AI infrastructure, including A5X systems based on NVIDIA’s Vera Rubin NVL72 and the eighth-generation TPU [1]. While Virgo currently manages connectivity within and across data center buildings, the company is already developing "Scale Across" networking, which aims to extend these capabilities via long-haul RDMA (Remote Direct Memory Access) to link accelerator clusters separated by hundreds of kilometers [1].

This evolution reflects a broader trend where networking is no longer treated as generic infrastructure but as a component co-designed with specific generations of accelerator hardware [1]. By separating the back-end fabric from the front-end, Google intends to evolve its AI networking independently of general-purpose compute resources, potentially allowing for more rapid upgrades as new GPU and TPU technologies emerge [1].

What to watch

• Multi-vendor interoperability: Monitor how Google integrates Virgo with non-proprietary equipment, as the company has stated a focus on protocol unification for distributed fabrics [1].
• Long-haul performance: Watch for updates on the "Scale Across" initiative, specifically how Google manages the latency and congestion challenges of linking geographically dispersed AI clusters [1].
• Deployment scale: Track the rollout of Virgo across Google Cloud’s global regions as it replaces or augments existing fabric for the A5X and TPU 8t infrastructure [1].

The success of Virgo will likely hinge on whether its "invisible" failure model can maintain consistent performance during the massive, multi-site training runs that define the current AI arms race. As Google pushes toward architectures supporting nearly one million GPUs, the ability to manage network-level bottlenecks will be as critical to performance as the raw power of the underlying silicon [1].