Methodology -- CloudBench

Data Sources

CloudBench aggregates performance data from multiple sources to provide a comprehensive view of Google Cloud Compute Engine capabilities.

GCP Official Documentation

Machine type specifications, network bandwidth limits, storage IOPS/throughput specs, and GPU configurations. This is our primary source for hardware specifications and theoretical maximums.

Geekbench Browser

Single-core and multi-core CPU benchmarks from publicly submitted Geekbench 6 results running on GCP instances. We use median values from multiple submissions for each machine type.

GCP Performance Benchmarks

Google's own published benchmark data for memory bandwidth (STREAM), network throughput, and storage performance across VM families.

Community Benchmarks

Real-world workload benchmarks from published blog posts, research papers, and open-source benchmark suites. Includes Nginx, PostgreSQL, Redis, FFmpeg, LINPACK, and ML inference results.

GPU Vendor Specifications

NVIDIA published specifications for H100 and L4 GPUs, including FP8/FP16 FLOPS, memory bandwidth, and interconnect speeds.

Machine Configurations

Unless otherwise noted, benchmarks use these standard configurations:

Parameter	Value
Region	us-central1 (Iowa)
OS	Ubuntu 22.04 LTS (or latest supported)
Boot Disk	pd-ssd (balanced persistent)
Local SSD tests	NVMe interface, maximum supported count
CPU benchmarks	Standard machine type with default vCPU count
Pricing	On-demand rates, us-central1, Q2 2026

Workload Benchmark Methodology

Real-world workload benchmarks follow standardized configurations to ensure fair comparison across VM families.

Web Serving

Nginx static file serving with wrk benchmark tool. 100 concurrent connections, 10-second test duration, measuring sustained requests/second. Node.js tests use Express.js with JSON serialization.

Tools: wrk, ab (Apache Bench)

Database

PostgreSQL tested with pgbench using TPC-B workload profile. 32 concurrent clients, 60-second runs. Redis tested with redis-benchmark using GET/SET mix with 50 parallel connections.

Tools: pgbench, redis-benchmark

ML Inference

LLM inference tested with vLLM serving framework. Image classification with PyTorch and standard batch sizes. All GPU tests use FP16 precision unless noted otherwise.

Tools: vLLM, PyTorch, llama.cpp (CPU)

Video Encoding

FFmpeg H.265 encoding from 1080p source material using libx265 "medium" preset for CPU tests. GPU tests use NVENC hardware encoder. Measured in frames per second.

Tools: FFmpeg, libx265, NVENC

HPC

LINPACK (HPL) benchmark for floating-point performance with vendor-optimized BLAS libraries. OpenFOAM CFD simulation using the standard motorbike tutorial case with OpenMPI.

Tools: HPL, Intel MKL, AMD BLIS, OpenFOAM, OpenMPI

Caveats and Limitations

1.Performance varies. Cloud VM performance can vary by 5-15% between instances due to hardware lottery, noisy neighbors, and thermal conditions. Numbers here represent typical/median performance.
2.Curated, not live. This data is sourced from published benchmarks and documentation, not from live testing. Your actual results will depend on your specific workload, configuration, and timing.
3.Pricing changes. Google Cloud pricing is subject to change. Always verify current pricing through the GCP Pricing Calculator for production decisions.
4.Not all metrics are equal. Synthetic benchmarks (Geekbench, STREAM) measure isolated capabilities. Real-world performance depends on the interplay of CPU, memory, storage, and network -- which is why we include workload-specific benchmarks.
5.Single-cloud focus. CloudBench exclusively covers Google Cloud Platform. We do not provide multi-cloud comparisons against AWS or Azure.

Data Freshness

Benchmark data was last curated in May 2026. We aim to refresh data quarterly as new VM series are released and pricing changes. If you notice outdated information, please let us know.