AI For Dev

methodology

3 min read

Benchmark Methodology

These pages are historical capability snapshots of every personal machine I’ve run Geekbench AI on. They are not a substitute for model-specific tests like Gemma 4 VRAM Requirements — they answer a different question: “How does this physical rig stack up against the others I own?”

Rig identity

A rig is identified by its CPU + motherboard/chassis combo. That pair is the only thing that doesn’t change as I swap GPUs, RAM, and drives over time. A computer name like “MSI Codex R” is just a friendly label for that pair.

Each rig page lists every CPU, GPU, and NPU run executed on that hardware, regardless of when. Swapping a GPU produces a new row in the same rig’s table, not a new rig.

What Geekbench AI measures

Three precisions across ten real-world AI workloads (Style Transfer, Object Detection, Pose Estimation, Image Super-Resolution, Image Classification, Face Detection, Depth Estimation, Text Classification, Machine Translation, Image Segmentation):

  • Single Precision (FP32) — full precision math. Headline number. GPUs almost always win here.
  • Half Precision (FP16) — what most modern inference actually uses. Apple Neural Engines and NVIDIA tensor cores shine.
  • Quantized (INT8) — what local LLM serving actually uses. NPUs and Apple Neural Engines often beat dedicated GPUs.

Each backend (ONNX/DirectML, CoreML/GPU, CoreML/NeuralEngine, OpenVINO/CPU, OpenVINO/NPU) gets its own run.

AI Capability Index (ACI)

A single comparable score per rig. For each rig, take the best Single, Half, and Quantized score across any backend tested on that machine, then:

ACI = (BestSingle + BestHalf × 0.6 + BestQuant × 0.4) / 100

Why these weights?

  • 1.0 on Single — the canonical Geekbench AI headline. Reflects raw FP32 throughput.
  • 0.6 on Half — FP16 is what most production inference uses today. Weighted lower than Single only because Single is the apples-to-apples baseline across all backends.
  • 0.4 on Quantized — INT8 is critical for local LLMs but is less universal (some backends fake it via FP32 fallback). Lower weight prevents Apple NE quant scores from overwhelming everything.
  • ÷100 — gives a clean two-to-three-digit number.

“Best across any backend” is intentional: a rig’s capability is what its best component can do for a given precision. If you want to compare individual backends, use the per-row scores in the rig page.

Tiers

TierACI RangeWhat it means
S700+Workstation-class. Comfortable with 30B+ local models.
A500–699Strong creator/AI rig. 13–27B models comfortable, 30B at Q4–Q5.
B300–499Capable laptop / older desktop. 7–13B comfortable.
C150–299Light AI work. Sub-7B models, vision tasks.
D<150Hobby / CPU-only territory.

What ACI does not tell you

  • VRAM ceiling — a rig’s ACI is irrelevant if a model doesn’t fit. Use Gemma 4 VRAM Requirements for that.
  • Token/sec for LLMs — Geekbench AI is computer-vision heavy. LLM throughput depends on memory bandwidth, not compute.
  • Thermals over time — these are short benchmarks. Sustained workloads tell a different story.
  • Cost — paired separately as $/ACI once pricing is filled in.

Pricing notes

Build cost is what I paid when I bought the parts, not current market value.

What’s tracked:

  • Whole-system prices for prebuilts and laptops
  • Individual part prices for DIY/swap builds: motherboard, CPU, RAM, GPU
  • Shared physical parts (one kit, multiple machines) are listed on every rig page they appear on but counted once in totals

What’s not tracked:

  • Cases, PSUs, coolers (granular tracking is impractical for piecemeal builds)
  • Storage drives
  • Cables, fans, accessories

See Also

Graph View

Backlinks