Skip to content
Cloud Provider Benchmarks

Benchmarks

Cloud compute benchmarks: CPU, memory and VM price-performance

Compute benchmarks compare virtual machines with the same vCPU and memory size across providers. Each run records the instance type, region, CPU model reported by the guest, kernel and tool versions, so results can be reproduced and compared over time.

Updated:

Direct answer

Compare cloud VMs on equal vCPU and RAM sizes, measure single- and multi-thread CPU throughput, memory bandwidth and boot time, then divide by the published hourly price. Each result shown here is recorded with its source, method and run metadata; this page also documents the method and the published price data.

What is measured

  • Single-thread and multi-thread CPU throughput (events per second)
  • Memory bandwidth (MB/s) for read, write and copy patterns
  • Time from API create call to SSH-ready instance
  • Hourly list price normalised per vCPU and per GiB of memory

Tools

  • sysbench — CPU and memory micro-benchmarks
  • stress-ng — Sustained load and throttling checks

Provider capability matrix

Every value is based on provider documentation. Undocumented features are shown as unverified.

Documented capabilities
Criterion PortvMind AWS Azure Google Cloud DigitalOcean Vultr Hetzner
Compute & GPU
Virtual machine instances Self-service virtual machines with selectable CPU/RAM profiles. Yes Yes Yes Yes Yes Yes Yes
Per-second or per-minute billing Standard compute billed at per-second or per-minute granularity rather than per started hour. Yes Per-minute billing. Yes Per-second for Linux on-demand (60 s minimum). Unverified Yes Per-second with a one-minute minimum. Yes Per-second (60 s / $0.01 minimum). No Hourly billing. No Per started hour with a monthly cap.
Dedicated / single-tenant resources Dedicated hosts, bare metal or organisation-dedicated cloud capacity. Yes Yes Dedicated Hosts and bare-metal instances. Yes Azure Dedicated Host. Yes Sole-tenant nodes. Partial Dedicated CPU Droplets; no single-tenant hosts. Yes Bare Metal. Yes Dedicated root servers.
NVIDIA H100 instances On-demand or reservable instances with NVIDIA H100 GPUs. Yes Published prices for 1–4 GPU profiles. Yes Yes Yes Yes Yes No
NVIDIA H200 instances On-demand or reservable instances with NVIDIA H200 GPUs. Yes Listed option; price not in public table. Yes Yes Yes Yes Unverified No
NVIDIA L40S instances On-demand or reservable instances with NVIDIA L40S GPUs. Yes Listed option; price not in public table. Yes No Azure Local only. No Yes Yes No
Platform & automation
API and Terraform automation Documented API plus an available Terraform provider for infrastructure as code. Yes Yes Yes Yes Yes Yes Yes
Open-source cloud stack (OpenStack-based) The provider publicly states that its IaaS control plane is built on OpenStack or an equivalent open-source stack. Yes OpenStack, Kubernetes and S3-compatible storage. No No No No No No

Metric definitions

Metric definitions
Metric Unit Better Definition
CPU single-thread events/s Higher sysbench cpu run with --threads=1.
CPU multi-thread events/s Higher sysbench cpu run with threads equal to vCPU count.
Memory bandwidth MB/s Higher sysbench memory, 1 MiB block, sequential write.
Provisioning time s Lower Create API call to first successful SSH login.

Verified results

Frequently asked questions

Why compare VMs on equal vCPU counts?

A vCPU is a scheduling unit, not a fixed amount of performance. Equal sizes make per-vCPU throughput and per-vCPU price directly comparable.