Home/All charts/Phone CPU Single-Core Efficiency
SPEC CPU 2026 · Teardown-tested · 20 Chips · 3 core types

Geekerwan Phone CPU Single-Core Efficiency Curves

Geekerwan phone CPU single-core efficiency curves cover 20 chips. In a simulated light-load scenario, the three phone CPUs scoring highest at the same board power are XRing O3, Tensor G6, Snapdragon 8 Elite Gen 5; ignoring power and looking at the ceiling under heavy load, the strongest peak belongs to A19 Pro.

Updated 2026-11-09View source data ↗
PEAK · Prime core · Float10.55A19 Pro @ 8.73 W

Geekerwan Phone CPU Single-Core Efficiency Curves

Mid core · Integer26 curves · axis 0–3.9W
Performance (points)Board power (W)0123401231W
How to read this chart
Solid / dashedIf one chip has two curves in a view (a few mid cores have two frequency variants), the solid line is the main variant and the dashed one the second.
Vertical dashed lineThe same-power comparison point of the current view — mid 1W, big 2W, prime 3W, chosen near each core type’s working range. The two tables below are fixed to the mid·integer 1W and prime·float peak scenarios; they do not follow this line. On the default view the 1W line happens to coincide.
Very short curvesSome models (both Apple and Huawei in this data) only publish 2–3 frequency points. Only those measured points are drawn, with no line between them, because the gaps were never measured. Models whose range does not cover 1W show “—” in the low-power ranking but still count in the high-power ranking.
Per wattPeak score ÷ peak power — how much score one watt buys at the peak point. Click the headers of the high-power ranking to sort.
Hover to readOnly the measured dots drawn separately on the chart can be read: they show the power and score actually tested. Continuous curves are fitted upstream and have no per-point measurements, so hovering them shows no reading — a fitted value is never presented as measured.
Cores should chain togetherOn one chip, the prime, big and mid cores should form a continuous chain — mid for light loads, big for medium-heavy, prime for the heaviest — so every workload has data. A gap means some loads have no curve.
Axes rescale per viewEach view scales its own axes, so the x-axis maximum changes with the view (see the range above the chart). A curve that looks “higher” after switching does not mean faster — read the ticks.

The two tables below are derived from the upstream dataset (equal-power scores are interpolated along the upstream curves) and provided for reference only. If you believe something is off, email us via the address on the About page.

Simulated low-power efficiency ranking

Mid core · Integer12 of 20 ranked1W board power · higher is better
#Chip1W
01XRing O3Xiaomi2.95
02Tensor G6Google2.46
03Snapdragon 8 Elite Gen 5Qualcomm2.45
04XRing O1Xiaomi2.45
05Dimensity 9600 ProMediaTek2.38
06Dimensity 9500MediaTek2.34
07Dimensity 9400MediaTek2.21
08Snapdragon 8 EliteQualcomm2.16
09Snapdragon 8 Gen 3Qualcomm2.12
10SnapDragon 8 Elite Extreme Gen 6Qualcomm1.91
11Snapdragon 8 Gen 2Qualcomm1.84
12Snapdragon 8+ Gen 1Qualcomm1.74
—Kirin 9050 ProHuawei—
—Kirin 9030 ProHuawei—
—A20 ProApple—
—A19 ProApple—
—A18 ProApple—
—A17 ProApple—
—A16 BionicApple—
—A15 BionicApple—

Simulated high-power efficiency ranking

Prime core · Float14 chipsDefault: peak score · headers switch to per wattPeak ↓
#ChipPeak power
01A19 ProApple10.558.73 W1.21
02SnapDragon 8 Elite Extreme Gen 6Qualcomm10.4215.53 W0.67
03Dimensity 9600 ProMediaTek10.3711.73 W0.88
04XRing O3Xiaomi10.3113.05 W0.79
05Dimensity 9500MediaTek9.4012.07 W0.78
06A18 ProApple9.337.84 W1.19
07Snapdragon 8 Elite Gen 5Qualcomm9.139.59 W0.95
08A17 ProApple8.356.48 W1.29
09Snapdragon 8 EliteQualcomm8.318.34 W1.00
10XRing O1Xiaomi8.238.12 W1.01
11Tensor G6Google8.0710.40 W0.78
12Dimensity 9400MediaTek7.839.19 W0.85
13A16 BionicApple7.315.28 W1.38
14A15 BionicApple6.704.92 W1.36

How to read these curves

What the two axes mean

The x-axis is “display-off board power”: the teardown runs on a dummy battery with the screen off, so idle board draw is only 0.05–0.1W and the number is essentially compute power (compute units and DRAM dominate). The y-axis is the SPEC CPU 2026 score. Each curve is one chip fitted from its lowest to its highest frequency: more power means more performance, and more heat and battery drain; the further a curve sits to the upper left, the stronger the performance — and usually the better the efficiency — at equal power, and the end of the curve is that core’s peak.

Core type × workload

The two selectors on the left multiply: core type picks which cluster is doing the work, workload picks what it is doing. Mid core plus integer is roughly what feeds your timeline, chats and navigation; prime core plus float is rendering, physics simulation and heavy gaming. One chip produces a different curve in each combination, so never compare across them.

Cores should chain together

On the same chip, the prime, big and mid cores should form a continuous chain — mid for light loads, big for medium-heavy, prime for the heaviest — so every workload has data. A gap means some loads have no curve. Snapdragon 8 Elite Gen5 is a good example: its mid and prime curves meet, covering everything from light daily use to single-thread heavy loads.

Equal power vs peak

To compare, cut a vertical line at one wattage: the higher curve is stronger at that power. The two tables below fix one scenario each — the low-power ranking reads the 1W score, the peak ranking reads the ceiling. A high ceiling is not free: A19 Pro draws 8.73W for 10.55 points, while SnapDragon 8 Elite Extreme Gen 6 needs 15.53W for 10.42.

Integer vs float

SPEC CPU single-core has two suites: integer (INT) covers compilation, compression and video encoding; float (FP) covers scientific work and physics simulation. Phone workloads are mostly integer, so integer performance matters relatively more.

Glossary

Board power (W)
Whole-board power measured with a dummy battery and the screen off. The test environment is clean: idle board draw is only 0.05–0.1W, so this is essentially compute power (compute units and DRAM dominate). Not battery output, not charging power.
SPEC CPU 2026
An industrial CPU benchmark built with fixed LLVM flags and workloads close to real applications. Scores are relative: higher is faster, and different versions are not comparable.
Prime / big / mid cores
Three clusters in one chip with different clocks and caches: prime handles single-thread peaks and heavy gaming, big takes medium-heavy work, mid carries most of your daily power draw. Their curves should chain together to cover the full range from light to heavy loads.
Integer (INT) / float (FP)
Two workloads: integer covers compilation, compression and video encoding; float covers scientific work and physics simulation. Phone workloads are mostly integer, so integer matters relatively more. Float usually costs more power and scores higher, and the two sets are not interchangeable.
Peak score / peak power
The highest point at the right end of a curve and the power it took to get there. A high peak does not mean everyday efficiency — always read it together with power.
Sweet spot
Where the curve starts to steepen: another step in clock buys less score and costs more power. Read it from the bend of the curve in the current view.

Related videos

Benchmark videos from Geekerwan — the tests behind these charts.

Single-core efficiency FAQ

What is the difference between integer and float?

+

Integer (INT) covers compilation, compression and video encoding; float (FP) covers scientific work and physics simulation. Phone workloads are mostly integer, so integer performance matters relatively more. On the same chip float usually draws more and scores higher, and the two cannot be converted into each other.

Why look at several core curves for one chip?

+

The three clusters cover different workloads: mid for light daily loads, big for medium-heavy work, prime for single-thread peaks and heavy gaming. A good design makes the three curves chain together, covering the full range; a gap means some loads have no data. To pick: prime for single-thread scores and heavy gaming, mid for daily heat and battery drain.

Why do the curves look taller after switching?

+

Each view scales its axes from that view’s own data: mid cores stay within a few watts, prime cores under float reach well over 10W. A taller-looking curve does not mean faster — read the ticks and the numbers.

Why do some chips only have part of the curves?

+

Upstream only ran per-frequency tests on some clusters for some models; no test, no curve. A few mid-core entries ship two frequency variants and are drawn as a solid plus a dashed line in the same colour.

More Geekerwan charts

Same chips, a different lens.