Announcing LotofBench, an Open-Source Alternative to Geekbench

At Cloud Mercato we benchmark cloud instances for a living — our platform tracks more than 23,000 VM flavors and millions of prices, and every performance claim we publish starts with a measurement we can defend. There is one tool the industry quotes more than any other: Geekbench. It is a fine product, but it is a black box — closed source, closed score database, no way to audit what is actually being measured. Today we are fixing that for ourselves and for anyone who cares: LotofBench, our open-source, Geekbench-style CPU benchmark, is now available under the MIT license — with a public, integrity-checked results board.

What it is

LotofBench runs sixteen CPU workloads that mirror the Geekbench 6 list — file compression, navigation, HTML5 browser rendering, PDF rendering, photo library import, compilation, text processing, asset compression, object detection, background blur, horizon detection, photo filtering, HDR fusion, ray tracing, structure from motion and object removal. Every workload runs single-core and multi-core, and publishes a score computed as throughput / baseline × 1000 against a fixed, documented baseline.

The whole suite is written in C11 with zero external dependencies — no library to install, no interpreter, no container required — and builds from source with one command on Linux, macOS and BSD.

Everything is implemented from scratch inside the repo: an LZ4 codec, SHA-1, a regex engine, a Markdown-to-HTML converter, a PDF parser, a software rasterizer, a Whitted ray tracer with a BVH, a quantized MobileNet-shaped neural network — the works. That is not machismo; it is what makes the suite auditable end to end. When a score moves, you can read every line of code that produced it.

Deterministic, verified, or it does not score

Benchmark suites live or die by reproducibility, so we made verification a first-class citizen:

  • Deterministic inputs. Every workload generates its data from a fixed seed — the same corpus, the same road grid, the same photos, byte for byte, on every machine.
  • Verification inside the timed cycle. Every result is checked (SHA-1 digests, route validation, exact counters, differential references) while it is being produced. A wrong result aborts the run instead of being scored.
  • Medians, not lucky runs. Five runs per workload, the median counts. Machine-readable output via --json.
  • A real test suite. 210 tests, clean under AddressSanitizer and UBSan, including differential tests that compare each optimized core against an independent, naive reference implementation.

A public results board you can audit

Scores follow the Geekbench conventions: the Integer and Floating-Point section scores are the geometric means of their member workloads, and the composite weighs Integer 65% against Floating-Point 35%. A default run on a 12-thread development machine scores 594 single-core and 2,973 multi-core. And to see how the numbers read in practice, here is the multi-core comparison of our first cross-calibration project — three AWS Paris generations (Graviton, AMD, Intel) — embedded live from our Projector:

All of it is published on the public results board, and the methodology page documents every rule behind the numbers — the sixteen workloads, the full baseline table and the publication pipeline.

The board publishes aggregates, not cherry-picked runs: average ± standard deviation over all verified results, one row per cloud instance type for official runs (performed by our internal fleet with the same public binary) and one row per CPU model for community uploads. Every row carries a build badge (Official / Community), an environment badge detected at bench time (Bare metal / VM / Container) and the system profile — CPU model, cores, RAM, OS, architecture. The first wave covers latest-generation instances from about twenty providers — Amazon Web Services, Google Cloud, Microsoft Azure, Oracle Cloud, Alibaba Cloud, OVHcloud, Infomaniak, Scaleway, Hetzner, Exoscale, DigitalOcean, Vultr, UpCloud, Linode, IONOS, Kamatera, VEXXHOST, T-Systems Open Telekom Cloud and more — and every official row links to the instance's catalog page.

And because a public board invites gaming, nothing reaches it untouched: every submission passes an automated integrity pipeline — a signed session with a proof-of-work, a full server-side recomputation of every score from the raw throughputs, pinned work amounts for the default seed, credible-throughput ceilings and a minimum runtime floor. Submissions that fail are rejected, logged for audit and never published. Results that pass publish automatically — no account, no manual curation.

What it is not

LotofBench is not affiliated with Primate Labs, and its scores are not comparable with Geekbench scores — the baselines are our own calibration. What they are is comparable across machines, architectures and time: a baseline never moves after a workload lands, and if it ever does, that is a major version bump by policy.

Try it

No build step required: each release ships prebuilt binaries for linux-x86_64, linux-aarch64, macos-x86_64 and macos-arm64 as tar.gz archives with SHA-256 checksums — the Linux builds are fully static, the macOS ones link the system libcurl. Three commands and your machine is benchmarked:

curl -LO https://github.com/cloudmercato/lotofbench/releases/latest/download/lotofbench-linux-x86_64.tar.gz
tar xzf lotofbench-linux-x86_64.tar.gz
./lotofbench

By default the run uploads its result to the board, where it passes the integrity check before appearing as a community row. Add --no-upload to keep everything offline, and ./lotofbench --json for machine-readable output. Prefer building from source? git clone then make — still one command, still zero dependencies. What gets published is your scores and the system profile — never your hostname; submitter IP addresses are stored only as salted hashes, for abuse control.

On the roadmap: measuring Geekbench 6 workload-for-workload on the same machines, so every LotofBench score gets its Geekbench 6 equivalent — a first cross-calibration project on three AWS Paris generations (Graviton, AMD, Intel) is underway. Pull requests are welcome.

Repository: github.com/cloudmercato/lotofbench — MIT license, contributions welcome. Results board: lotofbench.cloud-mercato.com.