Support center

The user guide for how VRAM Check scores, ranks, and explains local AI hardware.

Use this page when you want the rules in plain language: how the benchmark works, what makes a run official, how classes are assigned, and what the most common support questions actually mean.

Trust modelVerified, provisional, compatibility, estimated

The site keeps these states separate so results and claims remain interpretable.

Shared benchmarkTinyLlama public profile

Real inference, strict passes, runtime validation, and release verification are part of the public trust story.

How it works

A benchmark report, not just a score dump.

VRAM Check is designed to tell you what happened, why it matters, and what to do next.

1. Detect the machine

The CLI reads the machine first so the benchmark stays anchored to the hardware you actually own.

2. Run the shared profile

The benchmark runs one fixed public profile with strict pass selection so verified runs stay comparable.

3. Classify trust and ranking state

Each run is labeled as verified, provisional, compatibility, or estimated so the site never blurs one state into another.

4. Publish an evidence-backed result

The results page turns throughput, timing, stability, and bottleneck analysis into a readable report.

Methodology

What the canonical benchmark is actually measuring.

Decode throughput is the primary score

Leaderboard ordering starts with decode throughput. Prefill throughput and TTFT are tie-breakers and supporting evidence.

Strict mode exists to reduce noise

VRAM Check runs multiple canonical passes, selects a representative pass, and measures variance so noisy runs do not contaminate rankings.

Release verification matters

Official releases publish a checksum and can be verified locally with vramcheck verify-release --file .\vramcheck-windows.exe.

Ranking rules

Leaderboard positions only count when the run earns them.

Verified GPU required

Runtime acceleration must be active and the execution profile must be verified GPU, not compatibility mode.

Stability gates required

Decode, prefill, and TTFT variance must stay within the active policy to make rankings meaningful.

Leaderboard is not the same as visibility

Compatibility and provisional runs still produce valuable result pages, but they do not claim official competitive rank.

Class system

Class tells you capability. Rank tells you competition.

VRAM Check keeps absolute capability separate from competitive position so the report stays easier to read.

Class

Class is the absolute capability band of the machine under the active public benchmark.

Current public scale: S, A,B, C, D,E.

Rank

Rank is competitive context inside the active leaderboard scope.

A machine can have a modest class and still place well in a narrow field, or a strong class and a weaker rank in a crowded frontier board.

Key capabilities

Where the product is strongest today.

Verified GPU benchmarking

Runs one shared benchmark profile with real inference, runtime detection, release verification, and a strict stability policy.

Compatibility guidance

Shows what your current build can run comfortably, what starts getting tight, and where an upgrade would materially change the result.

Leaderboard discipline

Separates verified leaderboard runs from compatibility and estimated states so competitive claims remain honest and defensible.