How We Research AI Hardware
What we compare, where each number comes from, and how we decide the order of picks.
Last updated:
Specifications come from the makers
Spec tables use the maker's own page: the chip maker for architecture, cores and TOPS, and the brand that sells the exact product for size, power and ports. We prefer the India page. We never take specs from retailer listings, search snippets or forum posts. If a maker does not publish a figure, the table says "Not stated" and we do not rank on it.
Performance comes from named benchmarks
Speed figures, such as tokens per second for local language models or images per minute for image generation, come from third-party sources. We name the source under every chart and link to it. A chart compares products from one source with the same workload and settings, so numbers from different test setups are never mixed. We rewrite and explain the results in our own words and do not copy tables or charts.
Some benchmark sources test the reference version of a card or board while the product we link is a partner model. Where that is the case, the caption says so.
Which sources do we use, and what happens when they disagree?
Specifications come from maker pages: the chip or memory maker for the part, and the brand that sells the exact product for the board. Benchmarks come from independent publications and public test suites that show their method, test system and date. The benchmarks page lists every benchmark source used in a chart, what it measures and when we last read it.
When two sources give different numbers, we do not average them. We use the source whose test setup matches the product being discussed, say which one we used, and note the difference when it matters. A source that does not show its method or date is not used for a chart. A single vendor's own claim is labelled as the vendor's claim.
A source is replaced when a better-documented one appears, when a page is withdrawn, or when a new hardware generation makes the old figures a poor guide. The date we last read each benchmark source is shown, and material changes to an article are listed at its foot.
We do not test products ourselves
We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name. Our value is in compiling many published sources, checking them against each other and explaining what they mean for a buyer in India.
How picks are ranked
A pick needs the core specifications for its category to be published by the maker (for a GPU: the chip, memory size, memory type and board power). We then rank on the figures that decide the use case, such as VRAM for running larger models, and on the published benchmarks. Each pick lists what it is best for and who should skip it. Brands cannot pay to be included.
How we judge value
We do not print product prices. Retail prices change many times a day and differ by seller. Instead we judge value inside budget bands and performance tiers: which product gives the most of the deciding spec in its tier, what the platform around it costs in power supply, cooling and board class, and how long driver and software support is likely to last. Prices change, so check current prices before buying.
Corrections
Specs and availability change. Each article carries a "Last modified" date, and confirmed fixes are logged on the corrections page. If you spot a mistake, tell us and read our editorial policy for how we correct it.