Methodology

Units

Internal arithmetic uses bytes. 1 GB = 1,000,000,000 bytes. 1 GiB = 1,073,741,824 bytes. Discrete GPU marketing “N GB” is treated as N GiB for capacity. That assumption is shown on each hardware page and can be overridden with a custom capacity plus an explicit unit.

Weights

weightBytes = totalParameters × effectiveBitsPerWeight / 8 is the theoretical identity. When a specific checkpoint’s selected files are known, those file bytes are used instead. The calculator never sums every quantization sitting in a repository. Download size is not peak GPU memory.

Mixture-of-experts models store total and active parameter counts. Full-resident estimates use the total. Expert offload is not implied.

KV cache

For uniform full-attention layers:

kvBytes = 2 × numLayers × numKVHeads × headDim × cachedTokens × concurrentSequences × bytesPerKVElement

KV heads are used, not query heads. Head dimension prefers an explicit config value. Context is cached tokens per sequence (prompt plus reserved generation). Concurrent sequences share one weight replica.

What is out of scope in this version

• MLA, hybrid attention, and state-space caches as complete KV estimates.
• Training, LoRA, image/video, or game VRAM.
• Tokens/s prediction.
• Treating two GPUs as one larger GPU, or adding host RAM on top of discrete VRAM, or adding host RAM on top of unified memory.

Scenario ranges (low/typical/high extras) are not statistical confidence intervals. There is no claimed “95% accuracy”.

Calculator version 0.1.0, dataset 2026-09-19.1, schema 1.