Method
How we measure
How LeCompute measures, sources, and separates what is verified, estimated and hypothetical. The methodology behind the published numbers, and the public dataset repository.
Three statuses for every number
Every number published on LeCompute belongs to one of these three categories. When it is not obvious, the article says so explicitly:
The value comes from a citable primary source (vendor datasheet, JEDEC spec, paper, official repository). The link appears in the "Sources and method" section at the end of the article.
The value is computed from primary sources, with a documented and reproducible formula. The calculation is spelled out in the body of the text.
A working order of magnitude, flagged as such. We resort to it when no public source settles the question, but the reasoning does not hold without a bound.
The implicit bar: "an embedded or systems engineer reads it, and believes it." If a claim does not pass that filter, it does not pass publication.
Vendor "peak" numbers
Vendor datasheets (NVIDIA, AMD, Intel) advertise TFLOPS and bandwidth figures under ideal conditions: a single kernel, maximum sparsity, a perfect batch. Those numbers bear no direct relation to the sustained throughput observed in production. We quote them when they help compare two architectures, but we say so every time, and we give the sustained measurement alongside whenever one exists.
The same goes for 2:4 sparsity: NVIDIA systematically quotes the sparse values, which double TFLOPS on paper but apply to almost no real inference workload. On LeCompute, comparisons use dense values by default, unless stated otherwise.
The datasets are public
All the figures from our own measurements are published in a dedicated GitHub repository:
→ github.com/Pluenet-Killian/lecompute-benchmarks
Datasets under CC BY 4.0, scripts under MIT. Every data row carries its snapshot date and its source URL, which is what keeps the values interpretable once the pricing pages have changed.
The first published dataset: gpu-cloud-prices-europe.csv,
the on-demand hourly prices of AI GPUs at Scaleway, OVHcloud, Nebius,
FluidStack and a few non-European providers for comparison. Updated
roughly once a month.
Currencies
Every monetary value shown on LeCompute is in euros. Prices originally in USD are converted at the indicative rate of 1 USD = 0.92 EUR (mid-2026), and the rate is noted in the "Sources and method" section of the article.
Reproductions and corrections
If you can reproduce a benchmark we mention and your result diverges from ours, that is valuable information, for you as much as for us. Two channels:
- Issue or PR on the lecompute-benchmarks repository: the best way to propose a data correction or a new dataset.
- Email to redaction@lecompute.fr: to report an error in an article or suggest a subject.
Any published correction triggers an explicit update of the article
concerned (the updatedDate field), with a change note
whenever the revision affects a conclusion.