Cloud CPU & Memory Benchmarks, 2026 — who really wins per vCPU, per GB, per dollar.

TL;DR — In 2026 the ARM crown moved: Google Axion edges out AWS Graviton4 on per-thread compute, while AMD EPYC "Turin" (Zen 5) tops raw x86 single-thread. Graviton4 still leads on memory bandwidth (~537 GB/s). And on price-performance, Oracle Cloud (OCI) quietly wins almost everywhere — cheapest hourly rates across Intel, AMD, and ARM. There is no single winner; the answer depends on whether you optimize for speed, value, or a specific workload.

The contenders and their silicon

By 2026 every hyperscaler ships three CPU tracks — Intel, AMD, and their own (or Ampere's) ARM. The current flagships:

Two architectural notes that drive most of the results below. First, AWS Graviton4 and Google Axion are both Neoverse V2 designs — wide cores aimed at per-thread performance — whereas Azure's Cobalt 100 uses the lighter Neoverse N2, tuned for efficiency and density. Second, OCI bills ARM by the physical core (OCPU) where the others bill per vCPU/thread, which quietly tilts price-performance math in OCI's favour. [1,2]

CPU: single-core and multi-core

The standard kit — Geekbench 6, sysbench CPU, CoreMark, 7-zip compression, OpenSSL, and multi-thread scaling — tells a consistent story across the independent runs.

• Best per-thread ARM: Google Axion. In the DoiT cross-cloud ARM comparison, Axion (C4A) beat Graviton4 in 5 of 7 tests, by margins from 9.85% to 47.79%, and delivers "EPYC Genoa-level performance per thread." [3]
• Best raw x86 single-thread: AMD EPYC Turin (Zen 5). The dev.to/dkechag 2026 VM series is blunt: Turin "is simply a tier above anything else" in single-thread, and AWS's c8a posts the strongest multi-threaded result. [4]
• Graviton4 still delivers up to 30% more performance than Graviton3, and Azure's Cobalt 100 lands between Graviton3 and Graviton4 — strong, but a half-step behind the V2 pair. [1,5]
• Ampere AmpereOne M (OCI's newer ARM) finally improves on the aging Altra but roughly matches Graviton3 per-core — its win is multi-thread density and price, not peak. [3]
• Intel Granite Rapids is "a solid step forward" with steadier performance than Emerald Rapids — competitive, no longer leading. [4]

Consistency, throttling, and steal

Beyond peak scores, the production-relevant traits: the custom-ARM parts (Graviton4, Axion, Cobalt) run on dedicated cores with near-zero CPU steal and very low run-to-run variance — a deliberate design goal of vertically integrated silicon. Granite Rapids notably improved consistency over Emerald Rapids. Burstable/shared families (anything t-class or fractional-OCPU) still show steal under contention; avoid them for steady throughput. [4]

Memory: bandwidth and latency

Memory is where Graviton4 reasserts itself. STREAM-class bandwidth figures, normalized across the cited specs:

Graviton4's 12-channel DDR5-5600 controller gives it a real edge: AWS reports up to 50% better performance on memory-intensive database workloads vs Graviton3, and per-core memory bandwidth on r8g (Graviton4) runs ~7.5% ahead of c7a (Genoa). For latency and NUMA, single-socket ARM parts avoid cross-socket penalties entirely; multi-socket x86 instances pay a NUMA-hop cost on random access that STREAM hides but lmbench latency tests expose. [6,7,8]

Real-world workloads

Synthetic scores predict, but workloads decide. Patterns from the cited suites and HPC runs:

DevOps & build

• Docker build / CI pipelines: latency-sensitive, single-thread-heavy (dependency resolution, linking). AMD Turin and the V2-ARM parts (Axion, Graviton4) finish fastest. On price-performance, Graviton4 wins CI fleets running per-job.
• Java / Maven / Gradle: Graviton4 delivered 40% higher throughput per dollar on Java microservices vs equivalent x86 Xeon (Principled Technologies, 2025). The JVM is exceptionally well-tuned for Neoverse now. [5]
• Go / Rust compilation: heavily parallel; scales with core count and memory bandwidth — favours high-core Turin and Graviton4 instances.
• Kubernetes nodes: density + price/perf matter most. Graviton4 and OCI Ampere give the most pods-per-dollar; near-zero steal keeps the kubelet predictable.

AI & data (CPU-side)

• LLM token/s (CPU inference): Axion held a slight edge over Graviton4 on llama3.1:8b and llama3.2:1b token/second in the aarch64.cloud tests. [2]
• Pandas / Postgres / Redis: bandwidth- and latency-bound — Graviton4's memory edge shows up directly in query and cache throughput.
• Caveat: serious AI training/inference runs on GPUs/accelerators, out of scope here. This section is CPU-only data prep and light inference.

Cost-performance: the OCI surprise

This is where the rankings invert. Raw speed belongs to Axion and Turin, but dollars belong to Oracle:

• A peer-reviewed arXiv study (4 vCPU / 16 GiB, Intel) found AWS on-demand at $0.25/hr vs OCI's $0.14/hr — AWS costs ~79% more for only ~1.9% less runtime. OCI had the best cost-performance across all three architectures. [9]
• The dev.to/dkechag 2026 series agrees: "Oracle pretty much tops the performance/price charts with both ARM and x86." [4]
• Commitment changes the winner: on 1-year reserved, GCP Turin matches Oracle; on 3-year reserved, Azure Cobalt 100 unexpectedly tops the charts; spot instances deliver roughly 2× the performance-per-dollar of a 3-year reservation. [4]

Why the chips perform the way they do

• ARM (Neoverse V2): wide out-of-order cores, big caches, single-socket designs, and — critically — vertical integration. AWS/Google co-design silicon, hypervisor (Nitro / Titanium), and OS, eliminating steal and shaving virtualization overhead. That's why Graviton4/Axion feel more consistent, not just fast.
• AMD EPYC Turin (Zen 5): highest IPC and clocks of the x86 field, up to 192 cores, 12-channel DDR5 — it wins single-thread and dense multi-thread, but at higher per-vCPU list price.
• Intel: Granite Rapids closed much of the gap and brought stability, but fewer memory channels (8 vs 12) cap its bandwidth ceiling.
• Memory architecture is the quiet decider: 12 channels of DDR5-5600 (Graviton4) vs 8 channels (Intel) is a structural bandwidth advantage no clock speed fixes — and it dominates databases, analytics, and HPC.
• Hypervisor overhead: AWS Nitro and GCP Titanium offload virtualization to dedicated hardware, so guest CPUs see near-bare-metal performance; this is part of why hyperscaler ARM beats the same Ampere core running elsewhere.

Verdict

The honest bottom line: ARM is now the default for new general-purpose workloads — Graviton4 and Axion beat x86 on price-performance for the majority of cloud apps, and migration friction has largely evaporated. Reach for AMD Turin when single-thread latency is king (compilers, some databases), and reach for OCI when the finance team is watching. Pick the architecture for the workload, then pick the provider for the price model you can commit to.

References

• [1] AWS Graviton4 vs Google Axion vs Azure Cobalt vs AmpereOne — FaceOfIT.
• [2] ARM chip benchmark test for hyperscale cloud providers — aarch64.cloud.
• [3] ARM Wrestling: Benchmarking the Latest Cloud ARM CPUs — DoiT Engineering.
• [4] Cloud VM benchmarks 2026: performance / price — dev.to / dkechag.
• [5] The Real Cost of Graviton4 vs Cobalt 100, 2026 — workload breakdown.
• [6] Graviton4 benchmarked vs AMD and Intel — Tom's Hardware.
• [7] Graviton4 leading HPC performance — Arm developer blog.
• [8] Google Axion C4A vs Graviton4 performance review — Phoronix.
• [9] Cost-Performance Evaluation of General Compute Instances: AWS, Azure, GCP, OCI — arXiv:2412.03037.

Synthesized from published 2025–2026 benchmarks, cross-checked June 2026. Figures are from cited third-party runs, not original measurements; pricing is list/on-demand and changes frequently — verify before procurement.