TechCompare LogoTechCompare
Back

GitHub Actions runner benchmarks

Compare every runner in one table. See CPU hardware, single thread performance, and combined disk, cache, and internet scores from TechCompare's own tests.

26 runners across 9 providersResults imported 2026-10-10Pricing checked 2026-10-10How we test and score

All runner results

Select a column to sort. Hover for measurements or select a runner for the full details. Swipe across the table to see every score.

GitHub Actions runner CPU models, single thread CPU scores, disk, cache, and internet scores, and value per USD compute cost. Missing tests have no score.
Cost/vCPU/minUSDMax vCPU

Apple M5 Max

Apple M5 Max

Apple M4 Pro

AMD EPYC (model hidden)

AMD EPYC 4585PX 16-Core

AMD EPYC 9R45

Apple M4 Pro

AMD EPYC 4565P 16-Core

AMD EPYC (model hidden)

AMD Ryzen 9 7950X3D 16-Core

AMD EPYC (model hidden)

AWS Graviton (gen undisclosed)

Apple M2 Pro

Microsoft Cobalt 100

AMD EPYC 4585PX 16-Core

AMD EPYC 9V74 80-Core

AMD EPYC 7R13

ARM64 (model undisclosed)

AMD EPYC 7763 64-Core

AmpereOne family (SKU undisclosed)

Cobalt 100

Ampere Altra Q80-30

AMD EPYC 7763 64-Core

AMD EPYC 7502 32-Core

AMD EPYC 7763 64-Core

Intel Core i7-8700B

Single thread CPU scores are CoreMark iterations/s. Disk and cache scores use composite MiB/s, and internet scores use composite Mbps. Higher is better. Cache uses each runner's best tested cache or artifact backend. Each score uses only that runner's measurements, so adding a faster runner doesn't change existing scores. Value columns divide each score by USD cost/vCPU/min. K means thousand, M means million. These ratios don't estimate a build's cost. See the formulas.

Internet tests use Sharktech Amsterdam or Cloudflare. Hover or open details for the endpoint, protocol, and available sample counts. Compare matching endpoints, protocols, and transfer limits for a direct network comparison.

$/vCPU/min divides the quoted minute rate by the billed vCPU count. Prices and value use USD. RunJob's EUR rate is converted at €1 = $1.1206 (ECB, 2026-10-09). Prices are compute list or overage rates before allowances and extras. RunJob is consumed CPU only. GitHub uses paid private standard rates, and Namespace uses overage. Maximum sizes describe published runner offerings.

How we test GitHub Actions runners

TechCompare runs its own independent benchmark suite. We publish open results as downloadable summaries with file checksums and document the parameters behind our reproducible tests below. CPU results include five samples with the median, observed range, and variation. All operating systems and architectures appear together in the table.

These tests measure isolated workloads rather than whole application builds. The imported files don't include a run timestamp or workflow run URL, so the date above describes the import and pricing check.

CPU and compiler

CoreMark uses one execution thread, a 2000-byte dataset, five samples, and a 30-second target per sample. Linux and Windows pin the process to one logical CPU. macOS is unpinned. The median, observed range, and population coefficient of variation are preserved. Compilers differ between platforms, so these are runner plus compiler measurements.

Disk I/O

Disk tests use one worker at queue depth one, 1 MiB sequential I/O, and 4 KiB random I/O against a 1024 MiB file. OS file-cache bypass is required and writes include the final flush. Each operation repeats for at least five seconds. Disk scores use verified sample medians. The updated BuildPulse ARM result is included using its three completed samples. Completed and requested sample counts stay in the details.

Cache and artifact transfers

Each backend uses three unique-key save and restore pairs with a 512 MiB random payload. Restore runs in a fresh job, verifies the exact hit and SHA-256, and includes action and adjacent step overhead. Archive compression uses one thread. A 120-second publication gap is excluded from transfer time. Namespace Artifacts is a separate artifact service, not a volume cache hit.

Internet transfers

Internet tests use one HTTP/1.1 stream for three downloads and three uploads, capped at 1024 MiB and 60 seconds per sample. Protocol 7 calculates Mbps from actual bytes and elapsed time, including timed transfers. Upload counters track bytes accepted by the client socket. Tests use Sharktech Amsterdam or a Cloudflare fallback. Compare matching endpoints, protocols, and limits. Older timeout speeds are estimates.

How the scores work

CPU score is the measured median CoreMark iterations per second. Disk, cache, and internet scores use fixed geometric means of each runner's own measurements. Every component has equal weight. The formulas have no ranking reference or upper score limit, so adding a faster runner doesn't change existing scores.

  • Disk: the fourth root of sequential read MiB/s × sequential write MiB/s × random read MiB/s × random write MiB/s. Random throughput is IOPS divided by 256 because each operation transfers 4 KiB. All four measurements have equal weight.
  • Cache: the square root of save MiB/s × restore MiB/s. Throughput is payload MiB divided by elapsed seconds. The default picks each runner's highest score across complete, verified cache and artifact backends. Save and restore always come from the same backend. Hover or open details to see which backend was selected. The selector also offers GitHub Actions cache, provider archive cache, or artifact service comparisons.
  • Internet: the square root of download Mbps × upload Mbps when both are available. With one recorded direction, its Mbps value is the score. HTTPS request time stays in the details and doesn\'t contribute to the score.

Disk and cache scores are composite MiB/s, and internet scores use Mbps. Scores combine different workloads when several components are available. Filtering, sorting, or adding new runners doesn't alter the calculation. Disk and cache require every component. Internet uses the directions with recorded throughput. Raw measurements remain available on hover and in the row details.

Value columns divide the measured score by the USD cost per vCPU per minute. CPU value uses the single thread score. It doesn't scale that score by runner size. Disk, cache, and internet value columns start hidden and can be enabled separately. K means thousand and M means million. Every provider is sorted together using its USD rate. These are ratios for comparing benchmark performance with a quoted compute rate. They don't measure the cost or runtime of a complete build.

Prices and value use USD. RunJob's EUR rate is converted at €1 = $1.1206 (ECB, 2026-10-09).

A provider's largest published runner is a capacity option, not a measured result. A vCPU can be a hardware thread or a physical core depending on the platform. The normalized USD price follows the billed size. Billing rounding, free minutes, monthly plans, cache storage, and actual job duration can change the final bill.

Raw files, checksums, and research sources

Saved summaries, imported 2026-10-10. Full JSON dataset.

Avrea raw report

SHA-256 3031b688a95d29fd1a7025da2417a10f8af62c23e7ba84fad987db8e378fc06a

Blacksmith raw report

SHA-256 0cb9116e1ee1c2ff168bc472c565d91cf40ba226e00a6f79f3f12365ada7eb04

BuildPulse raw report

SHA-256 4af093479b3978ade6030c6ba2057709c5e01945dc7eee6fd898e5e85b51a66d

GitHub raw report

SHA-256 39b8dd0dfc8e149c91cb832950eae941a858b569a1abdf1c61f1c8c6b248703a

Latchkey raw report

SHA-256 63222d40f2f965e0ecadac1d1c0fcb7cb7cef203a5dd34a82f8c064410ee6509

Namespace raw report

SHA-256 ca43c4613af627ff2a66fd864ba0b6c211191fb862b63e26f2ef1c220079eb36

RunJob raw report

SHA-256 ef3f59b06254e722d2fb7585698dbda482dcab981a91beecf93285ac1d4dcd95

StarSling raw report

SHA-256 04b4b622e4751ae0c0bbcdf890e1cd2ce91911717569ae87eba225a48547903a

Ubicloud raw report

SHA-256 7c8dac9f2b28225f3ca8236dd3788596d6aacba0df73b869f19002adcbc68c6c

Protocol 7 internet report

Separate network update for 16 runners, imported 2026-10-10. CPU and cache use the original reports. Disk updates are listed separately.

SHA-256 59ff94b843d772437b28c412e4d9b02bb17dc5d32b702e9f4587e5a57e01ec02

BuildPulse ARM disk report

Separate disk update, imported 2026-10-10. The three completed samples are included in the disk score.

SHA-256 b0a3d6d39bac3cd65b444d10a726e6fa6f788fe1e9e23c7502c9df0258e7faa3

File hashes identify the exact bytes in each downloadable report.

Avrea pricing and hardware ↗Avrea runner sizes ↗Blacksmith pricing ↗Blacksmith instance types ↗BuildPulse rates and sizes ↗BuildPulse configuration reference ↗GitHub runner rates ↗GitHub standard runner sizes ↗GitHub larger runner sizes ↗GitHub Linux ARM CPU ↗Intel i7-8700B specifications ↗Latchkey plans and runner rates ↗Namespace pricing and hardware ↗Namespace 64 vCPU availability ↗RunJob pricing and per-job limits ↗RunJob CPU billing ↗StarSling prices, sizes, and CPU family ↗Ubicloud current runner rates ↗Ubicloud runner sizes and CPUs ↗Ubicloud premium routing ↗

Provider information checked October 10, 2026. Namespace profile sizes were confirmed from the supplied account settings.

Runner rankings and guides

Compare performance and value, or take a closer look at each measured provider.

Runners with the best CPU
Current leaders and the full ranking
View details ➜
Runners with the best CPU value
Current leaders and the full ranking
View details ➜
Runners with the best disk
Current leaders and the full ranking
View details ➜

Frequently asked questions

Are these independent GitHub Actions runner benchmarks?
TechCompare publishes measurements from its own runner benchmark suite. The results come from the supplied published reports, while provider documentation supplies hardware mappings, prices, and available sizes. Scores from provider marketing pages aren't used as test results.
Does a CoreMark score predict my build time?
No. CoreMark is a CPU microbenchmark running one execution thread. A full build also depends on parallelism, memory, disk, dependencies, cache behavior, and queue time. This snapshot doesn't measure complete application builds or queue delays.
How are disk, cache, and internet scores calculated?
Each score is a geometric mean of that runner's own measurements. Disk combines sequential read/write MiB/s with random read/write IOPS converted to MiB/s using the 4 KiB operation size. Cache combines save and restore MiB/s from one backend. By default, each runner uses its highest verified score across tested cache and artifact backends. Hover or open details to see the backend, or select a specific backend type. Internet combines available download and upload Mbps, or uses the single recorded direction. Protocol 7 reports speeds from actual payload bytes and elapsed time, including transfers that reach the time limit. Older timeout records use recorded bytes divided by the nominal limit. These fixed formulas have no ranking reference or upper score limit, so adding a faster runner doesn't change existing scores. CPU keeps its measured CoreMark iterations per second.
How is price per minute per vCPU calculated?
A fixed runner's quoted minute rate is divided by its billed vCPU count. Prices are shown in USD and are list or overage rates before allowances and extras. RunJob publishes EUR consumed CPU rates, which are converted using the dated ECB reference rate shown on the page. Its original rate stays in the details, and memory charges are additional. GitHub uses private standard rates, and Namespace shows overage with prepaid equivalents.
What does CPU value mean?
CPU value is the measured single thread CoreMark score divided by the USD cost per vCPU per minute. More billed vCPUs don't multiply the measured score. EUR compute rates are converted before calculating value, so every provider is in one USD ranking. K means thousand and M means million. Optional disk, cache, and internet value columns use the same score divided by rate formula. These ratios don't measure full build performance or predict the final bill.
Why is an exact CPU model sometimes missing?
Some virtual machines expose only AMD EPYC or aarch64. The table uses a provider mapping where documentation identifies a family, labels that evidence, and leaves undisclosed model numbers unconfirmed. Similar benchmark scores aren't enough to identify a processor.
Are missing tests counted as zero?
No. Internet transfers with a nonzero recorded byte count can supply a speed even if they reach the time limit. A blocked request or a timeout with no recorded bytes stays unscored. Disk uses verified sample medians with all four throughput components. The updated BuildPulse ARM disk result is included using its three completed samples, with the original counts in the details. Cache requires all three verified save and restore pairs. The table shows a dash when no score is available. Hover or open the row details to inspect the measurements.

Related tools

Data Transfer Calculator

Estimate transfer times for files over USB, WiFi, Ethernet, and more.

Use tool ➜

RAM Latency Calculator

Convert DDR3/DDR4/DDR5 timings (CL, tRCD, tRP, tRAS) into true latency in nanoseconds.

Use tool ➜

Memory and Storage Throughput Visualizer

Visualize the massive speed difference between CPU cache, RAM, and storage.

Use tool ➜

LLM API Pricing Calculator

Compare API costs across major models (OpenAI, Anthropic, Google) with prompt caching.

Use tool ➜

Use this tool on your own site

Drop the GitHub Actions Runner Benchmarks into your blog, docs, or CMS with one iframe snippet. Free, no signup. See all embeddable tools