Every score on this site is built from published benchmark figures, nothing else. Sub-scores run 0 to 100 against the best chip of the same kind here (per-watt ratios on a square-root curve so a 4x spread stays readable). The overall score is one fixed weighted sum per chip kind; there are no per-scenario weightings, because nobody measured those workloads. Where a credited source has run SPEC CPU 2017 on a chip, its single-core sub-score uses that instead of Geekbench and Cinebench: SPEC runs for minutes, so it sees throttling the short tests finish before. Values marked ≈ are estimates until checked. Sources: @geekerwan.

• 25% Single-core (Speed of one core: responsiveness, games, most apps. SPEC CPU 2017 when available, else gb6s, cb24s)
• 25% Multi-core (All cores together: rendering, compiling, exporting. gb6m, cb24m)
• 20% Efficiency (Multi-core work done per watt of rated power. cb24m per watt)
• 10% iGPU (Integrated graphics: light games and desktop without a card. ts)
• 20% Gaming fps

SPEC CPU 2017 integer, one core

The industry reference: minutes-long integer workloads on a single core. The hardest CPU number to game.

SPEC CPU 2017 floating point, one core

Science, media and simulation kernels on one core; leans on memory bandwidth as much as the core.

Geekbench 6 single-core

Short bursts of everyday tasks. Rewards boost clock and memory latency; a chip can ace it before it warms up.

Geekbench 6 multi-core

The same bursts on every core. Scales only partly with core count, so big chips look closer to small ones than they are.

Cinebench 2024 single-core

One core rendering for minutes. Sustained clock matters more than the peak on the box.

Cinebench 2024 multi-core

Every core rendering for ten minutes. Cooling and power limits decide this as much as the silicon.

3DMark Time Spy graphics (iGPU)

Integrated graphics only. Irrelevant once a graphics card is fitted.

Gaming, 1080p average fps (relative)

Game-suite average with a top-end GPU so the CPU is the limit. Cache size shows up here and nowhere above.

• 40% CPU (App launch, browsing, photo processing. SPEC CPU 2017 when available, else gb6s, gb6m)
• 30% Gaming (GPU throughput in demanding 3D games. wle)
• 30% Battery (CPU work per watt of peak power: longer runtime. gb6m per watt)

SPEC CPU 2017 integer, one core

The industry reference: minutes-long integer workloads on a single core. The hardest CPU number to game.

SPEC CPU 2017 floating point, one core

Science, media and simulation kernels on one core; leans on memory bandwidth as much as the core.

AnTuTu v10

Blends CPU, GPU, memory and UI into one number. Phones detect it and lift clocks, so treat it as a rough guide.

Geekbench 6 single-core

Short bursts of everyday tasks. Rewards boost clock and memory latency; a chip can ace it before it warms up.

Geekbench 6 multi-core

The same bursts on every core. Scales only partly with core count, so big chips look closer to small ones than they are.

3DMark Wild Life Extreme

A 20-second GPU run. Says nothing about how far a phone throttles ten minutes into a real game.

• 50% Raster (Traditional rendering: most games at high frame rates. ts)
• 30% Ray tracing (Ray-traced lighting and reflections. speedway, portroyal)
• 20% Efficiency (Raster performance per watt of board power. ts per watt)

3DMark Time Spy graphics

DirectX 12 raster at 1440p. Tracks most games well; ignores upscaling and frame generation.

3DMark Speed Way (ray tracing)

DirectX 12 Ultimate with ray tracing and mesh shaders: the forward-looking test.

3DMark Port Royal (ray tracing)

Older, lighter ray-tracing test; favours cards with weak RT hardware.

Blender 4.x samples per minute

GPU render throughput. Backends differ (OptiX, HIP), so cross-vendor gaps are partly software.