Methodology: How We Estimate Memory and Speed

How PersonalComputer.com estimates LLM memory needs and theoretical speed, and where our specs and prices come from.

Memory estimate

Total = weights + KV cache + 1 GB overhead. Weights = parameters (billions) × bits per weight ÷ 8, in GB. Bits per weight: Q4_K_M ≈ 4.89, Q5_K_M ≈ 5.70, Q8_0 ≈ 8.50, FP16 = 16 (llama.cpp's published figures for Llama 3.1 8B; llama.cpp: quantize README (bits/weight table)). For gpt-oss at Q4 we use 4.25 bits as an approximation of MXFP4 (4-bit values plus a shared scale per block); this is our assumption, not an OpenAI figure.

KV cache = 2 × layers × KV heads × head dimension × 2 bytes × context, for models whose layout we record (Llama 3.x and Llama-based distills). For other models we use a generic 1.5 GB per 8k tokens and flag it.

Fit rules

GPU: fits if total ≤ VRAM − 0.5 GB. Unified memory: we assume 96 of 128 GB (75%) usable on Ryzen AI Max and Mac, 24/48 GB on 32/64 GB Macs, and 120 of 128 GB on DGX Spark. These are planning assumptions; actual limits depend on OS and settings.

Speed ceiling

Tokens/s ceiling = memory bandwidth ÷ bytes of active weights per token. Real speed is lower. We do not present ceilings as benchmarks.

Sources

Accessed 21 Oct 2026. Retail prices are dated listings and change often.