Cloud VM Benchmarks 2026: AMD Turin Dominates Everything

If you select cloud VMs by provider loyalty or inertia, the 2026 benchmark data is worth your time. Dimitrios Kechagias ran comprehensive CPU performance tests across 44 VM types from 7 cloud providers — AWS, GCP, Azure, Oracle, Akamai, DigitalOcean, and Hetzner — with results dominated by a single processor that was not on last year's chart: AMD EPYC Turin.

AMD Turin Dominates Single-Thread Performance

Turin's single-thread performance lead is unusually clear-cut. In DKbench tests — covering 19 Perl and C/XS benchmarks representing real server workloads — Turin instances sit a full tier above everything else. The effect is largest on AWS's C8a, which runs Turin with SMT disabled. Without hyperthreading, each vCPU maps to a full physical core. The result: the C8a dominates multi-thread benchmarks by a margin that makes second place look distant. In NGINX tests, the C8a nearly doubles the score of the previous-gen Genoa C7a. For CPU-bound workloads where raw throughput matters, the performance gap is significant enough to justify the higher on-demand price.

Intel Granite Rapids Fixes Stability Issues

Intel Granite Rapids makes a meaningful appearance, mainly by fixing the instability that plagued Emerald Rapids. GCP's Emerald Rapids instances showed significant variance between regions due to boost clock sensitivity and node contention. Granite Rapids delivers higher and more consistent scores — a reliability improvement even where the absolute ceiling sits well below Turin. If you are on Intel and predictable performance matters more than peak numbers, the upgrade is worth it.

ARM Processors Close the Gap

On the ARM side, Google Axion (c4a) is the clear leader, delivering EPYC Genoa-level single-thread performance and competitive multi-thread results. Azure's in-house Cobalt 100 slots between Graviton3 and Graviton4 and offers competitive pricing. The new Ampere AmpereOne M finally delivers a tangible step up from the aging Altra. All ARM options benefit structurally in multi-thread benchmarks: each vCPU is a full core, so adding vCPUs gives near-linear throughput gains without the SMT tax that affects most x86 instances.

Price-Performance and Spot Instances

Price/performance rankings differ from raw performance. Oracle Cloud remains the best on-demand value for dedicated instances, with pricing comparable to what AWS and GCP charge for 1-year reserved VMs. Hetzner wins for budget workloads, though its cheapest shared-CPU types are Europe-only and sometimes unavailable. For the major three providers, GCP's n4d (Turin) and c4a (Axion) offer the best 1-year reservation value. AWS is competitive only with the C8a — older instance generations are the worst performance-per-dollar on the chart.

Spot instances shift the calculus further. GCP and Azure offer discounts approaching 80% on select types. For fault-tolerant or batch workloads, spot pricing from either provider can match or beat Hetzner on-demand. The full dataset and methodology are public, including a Docker image that runs the entire benchmark suite so results can be validated on any hardware.

KEY POINTS:

- 44 VM types benchmarked across AWS, GCP, Azure, Oracle, Akamai, DO, Hetzner
- AMD EPYC Turin leads single-thread performance by a clear margin over all others
- AWS C8a (non-SMT Turin) dominates multi-thread; nearly doubles Genoa C7a on NGINX
- Intel Granite Rapids fixes Emerald Rapids instability; performance more consistent
- Google Axion leads ARM; Azure Cobalt 100 and AmpereOne M add competitive options
- Oracle Cloud best on-demand value; Hetzner best budget option
- Spot instances (GCP/Azure) can match Hetzner on-demand at up to 80% discount