About VoidRuns
VoidRuns is a free toolkit for developers, engineers, and technical creatives who need precise answers rather than vague approximations.
The tools on this site are built around actual formulas and real benchmark data — not simplified sliders that give you a round number with no mathematical backing. The LLM VRAM Calculator uses the correct KV cache formula accounting for GQA architectures. The Blender estimator uses normalized scores from Blender Open Data. The code analyzer implements actual static pattern matching, not keyword detection.
All calculations run entirely in your browser. There is no backend, no user accounts, and no telemetry beyond standard Cloudflare analytics (which logs page views but not inputs). Your code, GPU configuration, and model choices are never transmitted anywhere.
Methodology
LLM VRAM: Model weight calculation uses the standard formula (params × bits ÷ 8). KV cache uses the per-head formula (2 × layers × kv_heads × head_dim × ctx × bytes) with accurate architecture configs for each listed model family, including GQA configurations. Runtime overhead is 10% of model weights + 0.5 GB fixed, validated against llama.cpp benchmark measurements.
Blender Render: Scores are derived from Blender Open Data 4.x results, normalized to RTX 3090 = 1000. Render time is calculated as base_time × (samples/128) × (1000/score). Base times are calibrated against the three official benchmark scenes (Monster, Junkshop, Classroom) averaged across their relative difficulty weights.
Code Complexity: Pattern matching uses curated regex sets for each language targeting documented anti-patterns from MDN, Microsoft .NET performance guidelines, and the CPython performance FAQ. Complexity inference is heuristic (loop nesting depth) and should be treated as a code review prompt, not a formal proof.