TensorCore.si is an independent, educational site about Tensor Cores, GPUs and AI chips. This page explains where our information comes from, how we check it, how we use AI tools, and how we correct mistakes. It applies to every article, tool and dataset on the site.
Who is responsible
Every page is edited and fact-checked by our editor, Sami Ullah Khan, who is responsible for what we publish. You can reach the team at hi@tensorcore.si.
Where our information comes from
We use sources in this order of preference, and link the ones we used at the end of each article:
- Primary sources: vendor datasheets, architecture whitepapers, product pages and official documentation from NVIDIA, AMD, Google, AWS, Intel and others.
- Research papers: peer-reviewed papers and preprints from the people who built a technique, such as the FlashAttention and DeepSeek-V3 papers.
- Standard and reproducible benchmarks: MLPerf results verified by MLCommons, and public benchmarks whose method and software version are published, such as the llama.cpp community results.
- Reputable reporting: established technology publications, mainly for dates, announcements and context.
When a figure is a vendor's own claim, such as "10 times lower cost per token", we say so in the text. We do not present marketing claims as independent results.
How we report numbers
- Per chip: rack and server totals are divided down to one GPU or accelerator, and we say when we have done this.
- Dense, not sparse: throughput is given without 2:4 sparsity. Where vendors only publish sparse figures, we halve them and note it.
- Precision stated: every TFLOPS figure names its number format, such as FP16 or FP8, because the same chip gives very different numbers at each.
- Estimates labelled: calculated values, such as memory needs or tokens per second, are marked as estimates and the formula is shown. Our calculators assume a GPU reaches 70% of its memory bandwidth during token generation, which matches the measured range on our benchmarks page.
- Dated: specifications change. Pages carry the date they were last updated, and vendor revisions are noted in the text, as with NVIDIA's 2026 change to Rubin's memory bandwidth.
How we fact-check
Before publishing, the editor checks every specification and key fact against at least one primary source, re-runs every calculation, and tests every link and tool. Comparisons are checked in both directions so that ratios and "faster" or "slower" statements match the numbers in the tables. If we cannot verify a number, we leave it out or mark it as not published.
How we use AI tools
We use AI tools to help research topics, draft text, write code for our calculators and check our work. A human editor reviews every page, verifies its figures against the linked sources and decides what is published. We never use AI to invent data, benchmark results, quotes, sources or personal experience.
Updates and corrections
We update pages when new hardware launches, when vendors revise specifications, or when better sources appear. Significant changes are listed on our site updates page.
If you find an error, email hi@tensorcore.si with the page and, if possible, a source. We review every report, fix confirmed errors and note significant corrections on the updates page.
Independence
TensorCore.si is not affiliated with NVIDIA, AMD, Google, Intel, AWS, Groq or any other company we cover. Companies do not pay for coverage or review our articles before publication. If we ever publish sponsored content or affiliate links, they will be clearly labelled.
Our tools and data
Our calculators run entirely in your browser and show the formulas they use. The GPU specs database can be downloaded as a CSV and reused with credit to TensorCore.si. See the disclaimer for the limits of estimates.