VRAM estimator
Weight memory is deterministic from parameter count × bytes/param. KV-cache and activation overhead are labeled ESTIMATE — we refuse to invent tokens/s benchmarks.
Weights (sourced math)
35.0 GB
KV [ESTIMATE]
28.0 GB
Activations [ESTIMATE]
7.0 GB
Total ballpark
70.0 GB
Formulas
- weights_GB = params_B × 1e9 × bytes_per_param / 1e9
- bytes_per_param(int4) = 0.5
- kv_GB ≈ params_B × context × batch × 0.05 / 1024 [ESTIMATE]
- activations_GB ≈ weights_GB × (1.2 − 1) [ESTIMATE]
- total_GB ≈ weights + kv + activations
- Weight memory is a lower bound from parameter count × precision.
- KV-cache and activation terms are ESTIMATES — not MLPerf or vendor claims.
- Framework (vLLM, transformers), parallelism, and quantization scheme change reality.
- If a figure cannot be sourced, this tool shows estimates rather than inventing tokens/s.