Methodology & Data Integrity
This document transparently describes how OpenGPU Radar collects, validates, and publishes pricing telemetry and hardware benchmark data. Our goal is to provide verifiable, reproducible compute cost intelligence for AI infrastructure decisions.
1. Pricing Telemetry & Ingestion Pipeline
OpenGPU Radar aggregates pricing data from four primary sources, each with distinct collection cadences and validation methods.
Vast.ai — Live Auction Index
We poll the Vast.ai public API (console.vast.ai/api/v0/bundles/) every 15 minutes during active collection windows. The API returns real-time auction listings with GPU model, VRAM, internet speed, reliability score, and asking price. We calculate the minimum spot rate and median verified host price across all listings matching our target GPU models (RTX 4090, H100 SXM5, A100, L40S). Hosts with reliability scores below 95% are excluded from median calculations. Price anomalies beyond 3 standard deviations from the rolling 24-hour mean are flagged and manually reviewed before publication.
RunPod — Catalog Baseline
RunPod does not expose a public pricing API. We maintain curated baseline rates derived from the RunPod console pricing page, updated weekly. On-demand rates reflect the standard cloud pricing tier. Spot rates reflect the community cloud minimums. These baselines are validated against user-reported deployment receipts and cross-referenced with the RunPod documentation. When RunPod adjusts pricing, we update within 48 hours of the change being publicly visible.
Lambda Labs — Instance Type Registry
Lambda Labs publishes instance types and hourly rates via their cloud console. We track on-demand pricing for all GPU instance types (H100 SXM5, H200, A100 80GB, L40S, RTX 4090, RTX 6000 Ada). Lambda does not operate a spot market — all pricing is on-demand with volume discounts available for reserved capacity. Our published rates reflect the standard single-node on-demand pricing without enterprise discounts.
Spheron — Bare-Metal Telemetry
Spheron provides bare-metal GPU nodes with transparent hourly pricing. We track both spot and reserved (bare-metal) rates. Spheron pricing is publicly listed on their platform with no hidden egress fees. We validate rates against actual deployment costs reported by users running workloads on Spheron infrastructure. The 0% virtualization overhead claim is based on direct root-access benchmarks comparing GPU passthrough performance.
Pricing Math Formulas
Monthly Cost (reserved): hourly_rate x 720 hours x 0.85
Monthly Cost (on-demand): hourly_rate x 720 hours
Break-even hours: hours where (spot_rate x hours) = (reserved_rate x 720)
Savings %: ((on_demand - reserved) / on_demand) x 100
The 0.85 multiplier on reserved pricing reflects the approximate 15% discount providers offer for 1-month commitments vs hourly on-demand. This factor is standardized across all providers for fair comparison.
2. Hardware Benchmarks & Test Environment Standards
All benchmark claims on OpenGPU Radar reference the following standardized test environment. GPU-specific benchmark results are derived from this configuration to ensure reproducibility.
| Component | Specification |
|---|---|
| Operating System | Ubuntu 22.04 LTS / 24.04 LTS |
| NVIDIA Driver | Linux 550.54.14+ |
| CUDA Toolkit | 12.4 |
| cuDNN | 9.x |
| Inference Runtime | vLLM v0.6.x (PagedAttention v2, FlashAttention-3) |
| Training Framework | PyTorch 2.4 with CUDA 12.4 backend |
| Quantization | FP8 (W8A8), AWQ (4-bit), BF16 baseline |
| NCCL | 2.20.x with NVLink 4.0/5.0 topology detection |
| Networking | InfiniBand NDR 400 Gb/s (where available) |
| Storage | Local NVMe PCIe 4.0 (7,000 MB/s sequential read) |
vLLM Throughput Methodology: Tokens/second measurements are collected using vLLM v0.6.x with PagedAttention v2 enabled. We serve Llama 3.3 70B (or equivalent parameter-count model) at batch=1 for single-request latency, and batch=64 for throughput saturation tests. FlashAttention-3 is enabled on Hopper and Blackwell architectures. Results are reported as sustained throughput over a 60-second window after warm-up.
FP8 Precision Validation: FP8 (W8A8) quantization is validated by comparing perplexity deltas against BF16 baselines on the WikiText-2 dataset. Acceptable degradation is <1% perplexity increase. On Hopper GPUs, we use the native FP8 Tensor Core path. On non-Hopper GPUs, FP8 results are estimated from INT8 benchmarks with correction factors.
Tensor Parallelism Benchmarks: Multi-GPU scaling efficiency is measured using NCCL all-reduce bandwidth tests across NVLink 3.0 (600 GB/s), NVLink 4.0 (900 GB/s), and NVLink 5.0 (1.8 TB/s) topologies. Scaling efficiency is reported as (measured_bandwidth / theoretical_peak_bandwidth) x 100.
3. Limitations & Preemption Disclaimers
Spot Availability Is Dynamic
Spot pricing on Vast.ai and RunPod community cloud is auction-dependent and changes in real-time. Published rates represent a snapshot at the time of data collection. Actual availability may vary based on time of day, geographic region, GPU model demand, and operator inventory. We recommend verifying current rates directly with the provider before making purchasing decisions.
Spot Eviction & SIGTERM Handling
Spot instances on Vast.ai and RunPod may be reclaimed by the operator with minimal notice (typically 30 seconds SIGTERM). Long-running training jobs should implement checkpointing at regular intervals. For production inference with strict latency SLA, dedicated on-demand or reserved instances are recommended. We document eviction notice times for each provider in our comparison pages.
Network & Storage Latency Variance
Storage throughput numbers (e.g., "7,000 MB/s NVMe") represent sequential read benchmarks. Real-world I/O patterns for model loading, checkpoint saving, and dataset access will vary. Network latency between nodes depends on the interconnect fabric (InfiniBand vs Ethernet), physical distance, and network congestion. Our benchmark environment uses dedicated network segments; shared-tenant environments may exhibit higher latency.
Pricing Exclusions
Published pricing excludes: data transfer/egress fees (which vary by provider and region), storage costs (NVMe, S3-compatible object storage), support plan fees, and enterprise volume discounts. Some providers (e.g., CoreWeave, AWS) require annual commitments for the lowest rates. Our published rates reflect the standard single-node on-demand pricing tier unless otherwise noted.
Benchmark Estimation
Some throughput numbers are estimated from architectural specifications (TFLOPS, memory bandwidth) rather than direct measurement. Estimated values are clearly marked with "est." or "~" prefixes. Directly measured values are collected from the test environment described in Section 2. We encourage users to run their own benchmarks using the documented test environment for production-critical decisions.
4. Data Corrections & Updates
OpenGPU Radar publishes pricing data on a daily cron schedule. If you identify incorrect pricing or outdated benchmark data, please open an issue on our GitHub repository. We review corrections within 48 hours and update the affected pages. All historical pricing data is preserved in our public git history for full auditability.