Cloud VM Benchmarks 2026: AMD Turin Dominates, ARM Solutions Gain Ground

Cloud infrastructure performance continues to evolve rapidly, with the 2026 benchmarks revealing clear leaders and emerging contenders in the competitive virtual machine market. Dimitrios Kechagias has expanded his annual comparison to include 44 VM types across seven major providers, testing across multiple regions to capture performance variations that often go unreported in official marketing materials.

Key Findings at a Glance

The most striking development this year is AMD's EPYC Turin processor, which establishes a clear performance lead over all competitors. This new architecture dominates both single-threaded and multi-threaded benchmarks, often providing such significant performance gains that it remains competitive even at higher price points. Meanwhile, ARM-based solutions continue to mature, with Google's Axion and Azure's Cobalt 100 offering compelling alternatives to traditional x86 architectures.

The Benchmarking Approach

For this year's comparison, the author focused on 2x vCPU instances, representing the minimum scalable unit for meaningful comparison. Most AMD and Intel instances use Simultaneous Multi-Threading (SMT), meaning their vCPUs correspond to Hyper-Threads rather than full cores. This distinction becomes crucial when evaluating performance and value.

The testing methodology included:

• Custom DKbench suite covering common server workloads
• Geekbench 5 for comparison with historical data
• Phoronix benchmarks including 7-zip, nginx, OpenSSL, and FFmpeg
• Testing across multiple regions to capture performance variations

Single-Thread Performance: Turin Sets a New Standard

Single-threaded performance remains critical for many workloads that cannot be easily parallelized. In this year's benchmarks, AMD's EPYC Turin establishes a clear lead, outperforming all competitors by a significant margin. The AWS C8a instance, powered by Turin with SMT disabled, provides a full core per vCPU and delivers exceptional single-thread performance.

Intel's new Granite Rapids shows improvement over its predecessor Emerald Rapids, offering higher and more stable performance. However, it still trails behind AMD's latest offering. Google's Axion ARM processor impressively matches EPYC Genoa-level performance per thread, making it the top ARM contender in this category.

Multi-Thread Performance and Scalability

Scalability testing reveals important differences in how cloud providers configure their virtual machines. ARM systems and certain x86 configurations (like AWS's C8a/C7a and GCP's t2d) provide 100% scalability, as each vCPU corresponds to a full physical core. Traditional SMT-enabled systems show lower scalability, typically in the 60-80% range.

The multi-threaded performance results reinforce Turin's dominance, with the AWS C8a leading the charts. Google's Axion performs competitively, closely followed by Graviton4 and Azure's Cobalt 100. The benchmark results across different test suites (DKbench, Geekbench, 7zip, nginx, FFmpeg, and OpenSSL) show consistent patterns, validating the methodology.

Performance-Value Analysis by Pricing Model

On-Demand Instances

For on-demand pricing, Oracle Cloud Infrastructure and Hetzner lead the value charts, with Oracle's Turin instances matching the performance of Hetzner's dedicated offerings at a similar price point. GCP's n4d Turin instances provide excellent value, while Google's Axion and Azure's Cobalt 100 offer competitive ARM alternatives.

AWS traditionally appears as the most expensive option, though their Turin instances provide better value than their older-generation offerings. DigitalOcean, once a top value provider, has fallen behind due to an aging fleet of processors.

Reserved Instances

Reserved instances offer significant discounts for commitments of 1 or 3 years. For single-threaded performance, GCP's Turin instances match Oracle near the top of the value rankings. Azure provides good value with their Cobalt 100 and Genoa options.

For multi-threaded workloads, the number of physical cores per vCPU becomes crucial. Oracle's ARM instances maintain top positions, joined by Azure's Cobalt 100. AWS's Turin instances (with full cores per vCPU) offer competitive value, especially at longer commitment terms.

Spot/Preemptible Instances

Spot instances represent the most cost-effective way to leverage cloud infrastructure, offering substantial discounts for fault-tolerant workloads. Oracle's fixed 50% discount on spot instances makes their Turin offerings consistently top value. GCP and Azure offer deeper discounts on their Genoa and Cobalt 100 instances.

For multi-threaded workloads on spot instances, Azure's Cobalt 100 leads the charts, followed by Oracle's AmpereOne M and GCP's older t2d Milan instances. Even AWS's Turin instances make the top ten, demonstrating their competitive performance despite less aggressive spot pricing.

Provider-Specific Recommendations

Amazon Web Services

AWS offers the fastest compute VMs overall with the Turin-powered C8a instances. While traditionally more expensive, their C8 family provides decent value, especially with spot instances or reservations. The absence of Graviton5 in public preview means Google's Axion currently leads in ARM performance.

Google Cloud Platform

GCP's 4th generation ARM (Axion) and AMD (Turin) instances offer the best combination of performance and value. The n4d Turin instances provide the best balance for most workloads, while the c4a Axion instances perform slightly better for multi-threaded tasks. Spot instance value varies significantly by region.

Microsoft Azure

Azure's in-house Cobalt 100 ARM processor delivers competitive performance and value, closely following Google's offerings. Their Dpls_v6 instances represent good value for reserved instances, particularly for multi-threaded workloads. Azure's spot instances offer deep discounts, making them compelling for fault-tolerant workloads.

Oracle Cloud Infrastructure

Oracle remains the top choice for small projects, offering a generous free tier with a 4-core ARM VM. Their paid instances provide exceptional value, with pricing comparable to reserved instances from larger providers. The AmpereOne M A4 and Turin E6 represent the best value for ARM and x86 workloads respectively.

Akamai (Linode)

Linode's shared-CPU instances offer good value, though performance varies depending on the specific CPU assigned. Dedicated instances now specify CPU generation, with the G8 Turin offering the best performance among their dedicated options.

DigitalOcean

DigitalOcean has fallen behind due to an aging processor fleet, though they still provide decent value for basic workloads. Their simple pricing and upgrade process remain advantages, but performance lags behind competitors.

Hetzner

Hetzner offers exceptional value, particularly their shared-core instances, though availability is limited to European regions. The CPX22 (AMD Genoa) provides the best combination of performance and value among their offerings.

Key Takeaways

1. AMD Turin dominates: The new EPYC processor establishes a clear performance lead across benchmarks, often justifying its higher price point.

2. ARM solutions mature: Google Axion, Azure Cobalt 100, and Ampere AmpereOne M offer compelling alternatives to x86 architectures, particularly for specific workloads.

3. Pricing models matter significantly: Spot instances offer approximately twice the performance-per-dollar of reserved instances, making them essential for cost-sensitive workloads.

4. Regional variations affect performance: The expanded testing across regions revealed significant performance differences, particularly for Intel processors.

5. Core configuration affects value: Instances providing full cores per vCPU (like AWS's C8a) offer better multi-threaded performance than SMT-enabled configurations.

For organizations evaluating cloud infrastructure providers, these benchmarks demonstrate that both performance and value depend on multiple factors beyond raw specifications. Workload characteristics, pricing models, regional requirements, and provider-specific configurations all influence the optimal choice. The complete benchmark data, available in the original article, allows for more nuanced decision-making based on specific requirements.

To see the comments or leave new ones, visit the original article on DEV.to.