In thread performance, Arm-based instances are catching up with Intel and AMD.
In the autumn of 2024, Dimitrios Kechiagas, a developer at SpareRoom (the peer-to-peer room-sharing site), noted an observation after testing 30 VM references from 7 providers (Akamai, Amazon, DigitalOcean, Google, Hetzner, Microsoft and Oracle).
This comparison has recently been updated. With the same seven providers, but a larger set of VM references (44). The focus remains on CPU workloads (general usage). And on the 2 vCPU configurations, most Intel and AMD instances now have multithreading enabled by default.
The testing methodology
The benchmark composition – available as a Docker image – has evolved slightly. Geekbench 5 remains part of it. Phoronix’s OpenSSL tests and 7-zip tests are still included, but the Linux kernel compilation has been replaced by an nginx test. Dimitrios Kechiagas added FFmpeg transcoding. The core of the benchmark continues with its in-house suite DKbench. It brings together about twenty tests including:
- Embedding CSS on wiki pages
- Applying image filters (Gaussian noise, blur mask, fractals…)
- Calculating prime numbers
- Counting codons in a bacterial sequence
- Finding a constellation from a celestial position (equatorial coordinates for a given epoch)
Where possible, 2 GB of RAM is paired with each vCPU and a 30 GB SSD is added as a boot volume. In most cases, the OS is Debian 13 64-bit (sometimes Ubuntu 24.04, and Oracle Linux on OCI’s Arm VMs). For a few instance types that impose a minimum of 4 vCPUs which cannot be disabled, the results are extrapolated. To obtain representative intervals, two or three instances are created at different times, and in various regions if possible.
Here we present a portion of the results. Specifically, the cost-per-performance ratios, expressed as DKbench points per $. The providers are ranked by their maximum ratio across the best of their tested instances*.
On-demand pricing
The cloud region with the lowest cost is taken into account between the United States and Europe.
In monothread
| Provider | VM | Min. | Max. |
| Hetzner | CCX13 (AMD Milan, 8 GB RAM, 80 GB SSD) |
53.80 | 70.70 |
| Oracle | Standard.E6 (AMD Turin, 4/30) |
54.40 | 55.10 |
| Akamai | Linode 4GB (AMD Milan, 4/80) |
29.20 | 37.70 |
| n4d-2 (AMD Turin, 4/30) |
30.10 | 30.60 | |
| DigitalOcean | PremAMD 2/4 (AMD Rome, 4/80, shared core) |
23.60 | 27.90 |
| Microsoft | D2pls_v6 (Azure Cobalt 100, 4/32) |
21.90 | 22.50 |
| Amazon | c8a.large (AMD Turin, 4/30) |
19.40 | 19.90 |
Thanks to the AMD Turin chips, Oracle comes closer to Hetzner. AWS is overall the least favorable value.
In absolute terms, better ratios are found at Hetzner with the shared-core CXAX11 (Ampere Altra; 122.30 to 130.80 points per $), the CPX22 (AMD Genoa; 134.10 to 142.10) and the CX23 (often Intel Skylake; sometimes AMD Rome, in which case the ratio ranges from 144.20 to 175.70). But their limited availability does not make them true competitors, says Dimitrios Kechiagas, which is why they were not included in the results.
In multithread
| Provider | VM | Min. | Max. |
| Oracle | Standard.A4 (Ampere AmpereOne M, 4/30) |
89.40 | 90.23 |
| Hetzner | CCX13 | 63.47 | 84.84 |
| Akamai | Linode 4GB | 57.63 | 73.96 |
| DigitalOcean | PremAMD 2/4 | 44.61 | 55.61 |
| Microsoft | D2pls_v6 | 42.89 | 44.69 |
| c4a-2 (Google Axion, 4/30) |
42 | 42.09 | |
| Amazon | c8a.large | 38.72 | 39.26 |
In this regard, three Arm-based instances dominate, all from Oracle. The Standard.A4, as well as the Standard.A2 (79.33 to 80.54 DKbench points per $) and the Standard.A1 (69.57 to 71.54). The Turin-based instance most interesting (65.93 to 66.86) is also available on OCI. AWS is closer to Google’s (Axion) and Microsoft’s (Cobalt 100) Arm configurations than in the monothread case.
On certain x64 VMs, multithreading is disabled. In that case, 2 vCPU equals 2 physical cores. This is the case, for example, for Amazon’s c7a.
DigitalOcean’s PremAMD 2/4 VMs, with shared cores, use AMD Rome CPUs. With Akamai’s Linode 4GB, also with shared cores, the assigned processor type can vary. Most often, it is AMD Milan.
Pricing with 1 year of commitment
In monothread
| Provider | VM | Min. | Max. |
| Hetzner | CCX13 | 53.80 | 70.71 |
| Oracle | Standard.E6 | 54.38 | 55.14 |
| n4d-2 | 46.89 | 47.71 | |
| Microsoft | D2pls_v6 | 35.84 | 36.95 |
| Akamai | Linode 4G | 29.21 | 37.67 |
| Amazon | c8i.large (Intel Granite Rapids, 4/30) |
26.85 | 27.64 |
| DigitalOcean | PremAMD 2/4 | 23.64 | 27.86 |
Google’s discount effectively brings it to parity with Oracle when AMD Turin chips are used. Azure proves interesting in both the Cobalt 100 and AMD Genoa configurations.
In multithread
| Provider | VM | Min. | Max. |
| Microsoft | D2pls_v6 | 104.34 | 108.73 |
| c4a-2 | 90.22 | 90.41 | |
| Oracle | Standard.A4 | 89.40 | 90.23 |
| Amazon | c8a.large | 77.94 | 79.02 |
| Hetzner | CCX13 | 63.47 | 84.84 |
| Akamai | Linode 4GB | 57.63 | 73.96 |
| DigitalOcean | PremAMD 2/4 | 44.61 | 55.61 |
In multithread, the number of physical cores makes a difference. AWS is in the game, even if Microsoft performs best with a Cobalt 100 instance.
Spot pricing (preemptible VM)
Prices are the lowest observed in the United States in January 2026.
In monothread
| Provider | VM | Min. | Max. |
| c3d-4/2 (AMD Genoa, 4/30, extrapolated from 4 to 2 vCPU) |
111.46 | 114.89 | |
| Oracle | Standard.E6 | 104.23 | 105.68 |
| Microsoft | D2pls_v6 | 93.20 | 96.06 |
| Hetzner | CCX13 | 53.80 | 70.71 |
| Amazon | c8a.large | 54.09 | 55.53 |
| Akamai | Linode 4G | 29.21 | 37.67 |
| DigitalOcean | PremAMD 2/4 | 23.64 | 27.86 |
With a fixed 50% discount, Google remains consistently attractive (twofold cheaper than on-demand) and thus on par with Oracle when AMD Turin chips are used. Azure is still interesting in the Cobalt 100 and in Genoa configurations.
In multithread
| Provider | VM | Min. | Max. |
| Microsoft | D2pls_v6 | 182.83 | 190.52 |
| t2d-2 (AMD Milan, 8/30) |
138.55 | 153.54 | |
| Oracle | Standard.A4 | 167.79 | 169.35 |
| Amazon | c8a.large | 108.23 | 109.74 |
| Hetzner | CCX13 | 63.47 | 84.84 |
| Akamai | Linode 4GB | 57.63 | 73.96 |
| DigitalOcean | PremAMD 2/4 | 44.61 | 55.61 |
The two best instances are Arm-based processors (Cobalt 100 and AmpereOne M). Google’s t2d on AMD Milan had already stood out in fall 2024 and remains in third place.
Overall, Hetzner’s CCX13 instances exhibit the most variation in performance depending on the data center where they are created.
Microsoft fell behind Amazon and Google in making AMD Turin chips available (GA by end of January 2026). Intel Granite Rapids are still in preview.
The DigitalOcean fleet is aging. Yet the low prices—and stable across regions—remain attractive for workloads where the cost-performance ratio isn’t the main concern.
* 12 VMs tested at Google, 11 at Amazon, 6 at Microsoft, 5 at Hetzner, 4 at Oracle, 3 at Akamai as well as DigitalOcean.
