Methodology
Last updated: 2026-11-15
How We Benchmark & Verify Data
FindBestRig tracks local AI hardware, Total Cost of Ownership (TCO), and live deals across UK and US retailers. We don't pretend to keep a warehouse full of every GPU and Apple Silicon SKU on the market. Instead, we pull verified measurements from the open-source community and fill in the gaps with physics-based memory bandwidth math so you can compare rigs on a level playing field.
1. Where Our Performance Numbers Come From
Every token generation speed (tok/s) on the site falls into one of two clear buckets:
- Measured Community Benchmarks (No Tilde Prefix): Compiled and normalized from reproducible
llama-benchand Ollama test suites—mainly the officialggml-org/llama.cppcommunity scoreboards andmaluramichael/llm-benchmarks. - Bandwidth-Derived Estimates (~ Tilde Prefix): Brand-new hardware—like a freshly released unified-memory Mac Studio tier or a Ryzen AI Max+ Mini PC—often lacks public runs for every model size. When that happens, we calculate an estimate and flag it with a tilde (e.g.,
~10.9 tok/s).
2. Our Bandwidth Calculation Formula
Training and prompt ingestion are compute-heavy. Token generation (decode speed) on dense LLMs is a completely different bottleneck: memory bandwidth. Your system has to read the entire model from VRAM or Unified Memory just to spit out a single token.
For estimated speeds, we take a chip's rated memory bandwidth, divide it by the quantized model weight, and multiply by a conservative 0.72 efficiency factor. That 0.72 multiplier accounts for attention overhead, KV cache reads, and runtime scheduling (based on Recal's Local LLM Bandwidth Ceiling analysis):
$$\text{Estimated tok/s} = \frac{\text{Memory Bandwidth (GB/s)}}{\text{Quantized Model Weight (GB)}} \times 0.72$$
Standardized Baseline Assumptions
To keep comparisons apples-to-apples across every rig on the site, all metrics assume the following baseline:
| Parameter | FindBestRig Standard | Why It Matters |
|---|---|---|
| Quantization | 4-bit (Q4_K_M / Q4_0) | The default standard in Ollama, LM Studio, and llama.cpp, retaining ~99% of base model quality. |
| 70B Model Footprint | 40.5 GB | Needs at least 44 GB of usable VRAM or Unified Memory (40.5 GB weights + ~3.5 GB KV cache and OS overhead). Anything under 44 GB is marked incompatible for native 70B VRAM execution. |
| Metric Reported | Token Generation (tg128) | Tracks steady-state decode speed—how fast words actually stream onto your screen—rather than initial prompt processing (pp512). |
| Offload Policy | 100% GPU / Unified Memory | Assumes zero CPU/system-RAM layer splitting or SSD swapping, which tanks generation speeds by 5× to 20×. |
3. Why Real-World Speeds Vary (±15%)
Your local setup won't match our baseline to the decimal. Expect roughly ±15% variance depending on four factors:
- Inference Engine: Running native MLX (
mlx-lm) on Apple Silicon often yields 10–15% faster speeds thanllama.cppthanks to tighter Metal kernel optimization. In those cases, our0.72multiplier acts as a conservative floor. - Context Window Length: Long prompts (32K–128K context) bloat the KV cache. More data to read per token means generation gradually slows down as your conversation drags on.
- MoE vs. Dense Architectures: Mixture-of-Experts models like Qwen3-Coder-Next or GPT-OSS-120B only activate a fraction of their parameters per token, running 2× to 4× faster than dense models of the exact same total size.
- Thermal Envelope: Cramming high-wattage silicon into a compact Mini PC or thin laptop causes thermal throttling during sustained, multi-minute generations. Open-airflow desktop towers don't have that problem.
4. Pricing, Currency & Affiliate Transparency
- Live Deal Tracking: We track UK (GBP) and US (USD) pricing against manufacturer specs and live retail listings. Where retailer rules forbid automated price caching—such as Amazon Associates—we link straight to the live product page rather than showing stale cached prices.
- Value Metrics (£/GB and $/GB): Calculated by dividing the lowest current in-stock rig price by total usable VRAM or Unified Memory capacity (
totalVramGb). - Affiliate Disclosure: FindBestRig is reader-supported. Clicking retailer links may earn us a commission at zero extra cost to you. Affiliate relationships never touch a rig's hardware specs, bandwidth numbers, or
tok/smath.