Methodology
This page explains how compatibility data on this Site is compiled, how confidence ratings work, and what our compatibility verdicts do and do not mean.
We do not test hardware
We have not physically tested any of the GPU, desktop, or LLM combinations described on this Site. Every verdict is derived from published specifications and documented third-party sources, not from installing a card in a machine and observing whether it boots. This is a deliberate design choice: it lets us cover far more hardware combinations than a hands-on review could, at the cost of not catching unit-to-unit variation (see Section 4).
If you are looking for hands-on reviews or benchmark videos, this Site is not that — it is a structured reference built from specification documents.
1. Source hierarchy
Sources are recorded at two levels, and the two do not claim the same thing.
The evidence behind a verdict. Each item of evidence for a compatibility verdict is recorded with the document it came from, the date that document was last checked, and a confidence grade. This is what the evidence list and the grades shown with each verdict display, and it is the level at which a verdict's reasoning is cited.
The specifications on a machine or card record. The published values held on a record are cited by the list of documents that record was compiled from, shown in the source table at the foot of the page. Unless the published dataset records a source, a checked date and a grade for that individual field, such a citation means "this record was compiled from these documents", not "this exact number came from this exact document". We do not claim more than the data holds: where a figure decides a verdict, read the evidence and the grades shown with the verdict itself, and treat the record's citation list as the documents the record was built from.
The grades are the ones assigned in the compatibility database itself, so that a grade shown on a page means exactly what the grade in the data means. We prefer evidence higher in this list and fall back to lower rungs only when nothing better exists:
- The manufacturer's own specification sheet or manual — the published documentation for the machine or for the card, read directly (grade A).
- A derivation from official specifications and published standards — nobody states the combination, but the published numbers settle it: a card's board power against the supply's rating, a card's length against a stated clearance, the power a PCI Express slot supplies (grade B).
- Verification on physical hardware, by whoever produced the measurement (grade C). Note that this Site does not itself test hardware, as set out above; a grade C value is one whose cited source was produced by someone who put the parts together.
- A third-party measurement or article — a teardown, a review, an enthusiast site that measured or reported the figure (grade D).
- A user report — a forum or community post describing what someone did (grade E).
A definite verdict — Fits, Fits with conditions, or Does not fit, shown without an "Estimated" badge — is made only on grade A, B or C evidence. Grade D or E evidence on its own produces an estimate.
2. Confidence grades (A–E)
Every verdict on the Site displays the grades of the evidence behind it, so you can judge how much weight to give it. More than one grade means more than one kind of evidence:
| Grade | Meaning |
|---|---|
| A | Stated explicitly in the manufacturer's own specification sheet or manual |
| B | Derived logically from official specifications and published standards (nobody states the combination; the numbers do) |
| C | Verified on physical hardware |
| D | Third-party measurement or article |
| E | User report only |
Where a kind of source usually carries a particular grade, the grade recorded on the individual verdict wins over that default.
Verdicts resting on grade D or E evidence alone are marked as Estimated with a distinct dashed badge wherever they appear, so they are never mistaken for a verdict resting on an official specification, and they are never blended into the surrounding prose. Where a summary or list combines data points of different grades, we do not present the combination as stronger than its weakest input: the individual pages carry the grades, and every list links back to them.
3. What the compatibility verdicts mean
| Verdict | Meaning |
|---|---|
| Fits | The published specifications show no apparent conflict — slot type, card length, clearance and power delivery are each within what the documents state. It is a statement about the specifications, not a prediction about your unit |
| Fits with conditions | Compatible per specifications, but only if a specific condition is met (e.g., a PSU upgrade, a specific slot, or removal of a drive cage) — the condition is stated on the page |
| Does not fit | Published specifications indicate a physical or electrical conflict (e.g., insufficient PSU wattage, no available PCIe slot of the required type, or the card is too long for the case) |
| Unverified | We do not currently have enough sourced data to render a verdict |
A "Fits" verdict is a statement about published specifications. It carries no guarantee about your specific unit. See Section 4.
4. Why a "Fits" verdict can still be wrong for your unit
Manufacturers publish specifications for a model line, but individual units vary:
- Power supplies from the same model line can vary in real-world output tolerance, and OEM power supplies are frequently rated close to their actual limit rather than with generous headroom.
- Prior owners may have modified, repaired, or replaced internal components (including the power supply itself) without updating any visible label.
- BIOS/firmware versions affect PCIe lane negotiation and resizable BAR support in ways that are not always documented per revision.
- Case revisions within the same model name sometimes change physical clearances slightly.
For this reason, always verify compatibility with the manufacturer, or physically confirm clearances and power requirements yourself, before purchasing a component. Leave headroom on power supply capacity rather than relying on a specification's rated maximum.
5. LLM / VRAM sizing methodology
For local LLM compatibility, we estimate the VRAM required to run a given model at a given quantization level (e.g., Q4_K_M) including a working allowance for the key-value (KV) cache at a stated context length. These figures are calculated, not benchmarked on the specific GPU listed — actual usable context length and multi-tasking headroom will vary by inference engine, batch size, and other loaded processes. The figures the calculation turns on — the card's memory size, the size of the file that is distributed, and the model's attention geometry — each carry their own source, the date it was read and their own grade, and because the outcome is calculated rather than stated by a vendor it is never better than grade B and never better than the weakest of those figures: where one of them is grade D or E, that grade is shown on the row alongside B and the row names the figure it applies to (see Sections 1–2).
6. Reporting an error
If you find a data point that appears incorrect — a wrong spec, a broken source link, or a verdict that doesn't match your own research — please tell us at contact@vramtable.com with the page URL and what you found. We review and, where confirmed, correct reported errors; the lastmod date on the affected page is updated when we do (see Section 7).
7. Update frequency
There is no fixed schedule for revisiting existing pages. Data is updated when: (a) a manufacturer publishes a revised specification, (b) an error is reported and confirmed (Section 6), or (c) we add a new desktop, GPU, or LLM to the database. The "last modified" date shown on a page reflects only actual changes to that page's content — we do not refresh dates automatically as part of a routine site rebuild.