Google has announced Virgo Network, a new megascale data center fabric purpose-built for AI workloads. It uses a flat, two-layer non-blocking topology with high-radix switches to reduce network tiers and latency. The architecture separates east-west accelerator traffic (Virgo) from north-south storage/compute traffic (Jupiter), enabling independent evolution of each layer. Key specs include support for 134,000 TPU chips in a single fabric, up to 47 petabits/sec of non-blocking bisectional bandwidth, 4x bandwidth per accelerator over the previous generation, and 40% lower unloaded fabric latency. Reliability features include independent switching planes for fault isolation, sub-millisecond telemetry, and automated straggler and hang detection to maximize training job goodput.
Sort: