Major cloud providers do not publish internal GPU fabric bandwidth specifications

Cloud providers (AWS, GCP, Azure) do not publish the internal bandwidth of their GPU cluster fabrics, making it impossible to compare cloud vs on-premises performance directly.

Cloud providers consider their network infrastructure a competitive advantage and do not disclose internal specifications. AWS EFA documentation describes the adapter but not the underlying fabric bandwidth. GCP documentation describes the interconnect but not the per-GPU bandwidth. Azure documentation is similarly opaque.

Benchmark actual performance using NCCL tests rather than relying on published specifications. Use the cloud provider's own performance benchmarks as a reference point. For AWS, the p5.48xlarge instances with EFA provide the highest GPU fabric bandwidth, but the exact number is not published.

This is an honest negative - the information is genuinely not public. Any claim about cloud fabric bandwidth should be treated as a vendor claim unless backed by independent measurement. The lack of transparency makes it difficult to make informed decisions about cloud vs on-premises deployment.