1. The Core Principles
Cloud GPU Rating was established to bring empirical rigor to AI infrastructure selection. Our ratings are governed by three non-negotiable rules:
- Zero Sponsored Placement: Providers cannot pay for higher rankings, badge placements, or favorable reviews.
- Real-World Hardware Measurements: We measure actual sustained FLOPS, NCCL all-reduce bandwidth, and disk I/O, rather than relying solely on vendor marketing claims.
- Full TCO Transparency: We factor in ingress/egress bandwidth fees, local storage rates, minimum billing granularity, and commitment cliffs.
2. Evaluation Dimensions & Weighting
| Evaluation Pillar | Weight | Key Metrics |
|---|---|---|
| Hardware Performance | 30% | Sustained FP8/BF16 TFLOPS, thermal throttling resistance, HBM3e bandwidth |
| Interconnect & Fabric | 25% | InfiniBand Quantum-2 (3.2 Tbps), RoCE v2, multi-node NCCL all-reduce latency |
| Price & TCO Efficiency | 20% | Hourly on-demand rate, spot discount depth, free egress vs per-GB egress tax |
| Operational Agility | 15% | VM spin-up speed, Slurm / Kubernetes GPU operator integration, REST API quality |
| Reliability & Compliance | 10% | Uptime SLA, SOC 2 / ISO 27001 compliance, GDPR sovereign datacenter residency |
3. Verification Protocol for Providers
Every provider profile in our directory undergoes verification before receiving the Verified Compute Operator badge. Verification requires documentation of physical datacenter locations, direct billing evidence, and confirmed GPU inventory.
4. Submitting Benchmark Data
Researchers and cloud engineering teams can submit reproduction logs, Slurm job run outputs, or pricing corrections through our contact form or by emailing [email protected].