Cooling the A20 Pro with liquid nitrogen barely raised its Geekbench 6 scores. In a comparison chart published by the technical testing channel Geekerwan, the gain over room temperature is just 0.76% in single-core and 2.49% in multi-core. That is smaller than the gains for the previous-generation A19 Pro and rival chips, which suggests the A20 Pro can deliver high CPU performance under ordinary cooling. But reading this as an "iPhone that runs as if cooled by liquid nitrogen" would misinterpret what the experiment shows. The A20 Pro's progress becomes clearer when you set the comparison with heavily cooled previous-generation chips alongside memory and compute-unit improvements, and the design of the benchmark itself.
Liquid nitrogen added just 2.49% in multi-core
Geekerwan's iPhone 18 Pro performance analysis was published on September 16, 2026. Its comparison chart (reposted image) lines up Geekbench 6 CPU scores measured under two conditions: room temperature and liquid nitrogen cooling. The A20 Pro's single-core score rises from 4,732 to 4,768, and its multi-core score from 12,727 to 13,044.
| Chip | Single-core: room temp → liquid nitrogen | Gain | Multi-core: room temp → liquid nitrogen | Gain |
|---|---|---|---|---|
| A20 Pro | 4,732→4,768 | 0.76% | 12,727→13,044 | 2.49% |
| A19 Pro | 3,866→4,089 | 5.77% | 10,095→11,111 | 10.06% |
| Snapdragon 8 Elite Gen 5 | 3,613→3,795 | 5.04% | 10,711→12,063 | 12.62% |
| Dimensity 9500 | 3,359→3,666 | 9.14% | 9,776→11,014 | 12.66% |
Scores are in points. The gain was calculated from the values in Geekerwan's chart as (liquid-nitrogen score ÷ room-temperature score − 1) × 100, rounded to two decimal places. The chart does not list specific device models, the OS, or the exact Geekbench version. Because the room and cooling temperatures, the number of repeated runs, and measurement variance are also unknown, the figures do not represent product-wide averages or statistically significant differences.
For the other three chips, strong cooling lifts multi-core scores by roughly 10–13%. Within the same chart, the A20 Pro gains the least from cooling. Working backward, its room-temperature multi-core score reaches 97.57% of its liquid-nitrogen score. For the A19 Pro, the figure is 90.86%. Both are ratios of the published values (room temperature ÷ cooled × 100), not measures of cooling efficiency.
In this CPU test, the A20 Pro has little score left to recover through extra cooling. One reading is that it relies less on extreme cooling to reach high performance. However, the chart contains no clock-speed or power-consumption logs, so it is not possible to attribute the small difference solely to thermal performance.
A generational gap remains even against a cooled A19 Pro
In Geekerwan's chart, the A20 Pro's multi-core score is 26.07% higher than the A19 Pro's at room temperature and still 17.40% higher under liquid nitrogen cooling. The comparisons are 12,727 vs. 10,095 at room temperature and 13,044 vs. 11,111 when cooled.
This generational gap uses a different baseline from the cooling gains in the previous section. It is calculated as (A20 Pro ÷ A19 Pro − 1) × 100 under the same cooling condition. It can be recalculated from Geekerwan's comparison chart, published September 16, 2026, but that does not mean power settings or the test devices were standardized.
The gap narrows when the A19 Pro is also cooled heavily, but the ranking does not reverse. The published results at least do not fit the view that the A20 Pro's high score is explained only by better cooling. There is a generational gap that persists under cooling, and there is also a change that lets the chip get close to that performance at room temperature.
It would be premature, though, to subtract 17.40% from 26.07% and treat the difference as the contribution of thermal design. The two generations differ in their compute circuits and memory systems, and cooling can also change how clock speed and power are controlled. The gap under liquid nitrogen is also not an improvement in IPC, the number of instructions processed per clock cycle. The chart shows how the generational gap persists, but it is not an experiment that isolates the contribution of each improvement.
A package that changes the path heat takes
With the A20 Pro, Apple adopted its own package that places the silicon die and memory side by side. According to the official announcement, the design draws inspiration from Apple's M-series silicon, removes the memory from the chip's heat-dissipation path, and bonds the A20 Pro directly to a next-generation vapor chamber.
Heat produced by the chip ultimately has to be moved out of the device. If any point along the way conducts heat poorly, even a large cooling component cannot receive enough of it. The side-by-side layout is a change to what sits between the heat source and the cooling component, and it plays a different role from simply enlarging a heat-dissipation part.
The side receiving the heat has changed too. Apple says the new vapor chamber uses a new material and has three times the surface area of the iPhone 17 Pro's. It revises both the path that carries heat from the chip and the component that spreads it. The high room-temperature score is consistent with this design approach.
Still, three times the surface area cannot be read as three times the cooling capacity. Apple's claimed "up to 40%" improvement over the previous generation refers to sustained performance, whereas the 2.49% gain in this Geekbench test is the increase when the same A20 Pro is cooled further. The comparison targets and what is being measured both differ. The small 2.49% figure neither refutes nor proves Apple's 40% claim.
What 2nm and memory bandwidth each change
The A20 Pro is built on a 2nm process, and its 6-core CPU integrates a Neural Accelerator. Apple also says memory bandwidth is up 50% over the A19 Pro. Alongside the thermal improvements, these are worth considering because, even with the same number of CPU cores, they can change how fast processing proceeds and how data is handed off.
On the manufacturing side, TSMC explains the power-efficiency gains from GAA nanosheet transistors in its research abstract on the N2 2nm generation. GAA is a structure in which the gate surrounds the channel through which current flows, an evolution in how transistor current is controlled. This is technical background for understanding the 2nm generation; it does not mean TSMC's process-level performance targets can be directly multiplied into the A20 Pro's speed or battery life. The only manufacturing process Apple's announcement specifies is 2nm.
Whereas cooling removes heat that has already been generated, improvements in manufacturing and circuit design affect how much power is needed to process work. They can be used to complete the same work with less power, or to raise speed within an acceptable power range. When looking at the A20 Pro's room-temperature score, it is necessary to consider separately the side that removes heat and the compute side that generates it.
The 50% increase in memory bandwidth is also a separate improvement from CPU clock speed. Bandwidth is the amount of data that can be moved in a given time, and it is different from memory capacity. Even if compute units get faster, an app will not run faster if it spends a long time waiting for the data it needs. Wider bandwidth creates room to reduce that waiting time, but it does not speed up every workload by 50%, especially those not limited by reads and writes.
Geekbench 6 is affected by this as well. A technical document from Primate Labs explains that CPU test performance depends on various parts of the processor and memory. Its multi-core test is designed so that cores share one task rather than each running independent jobs, which also requires communication and coordination between cores. Multi-core score differences are therefore hard to explain by core count or clock speed alone.
The GPU and AI circuitry deserve separate evaluation. According to Apple, the 7-core GPU is up to 40% faster than the A19 Pro's, and the Neural Engine, with dual 16-core units for 32 cores in total, delivers twice the AI processing capability. The GPU's 40% here is a different metric from the 40% sustained-performance figure in the previous section. The CPU Geekbench score alone cannot verify either claim.
In actual AI apps, Core ML uses the CPU, GPU, and Neural Engine and optimizes processing while limiting memory use and power consumption. Even if the Neural Engine is strengthened, not all of an app's processing runs on that circuitry. How much faster on-device AI becomes needs to be confirmed through measurements that use the same model and inputs and fix where the work is executed.
Pauses in Geekbench, and the distance to real-world use
Geekbench 6 inserts pauses between individual workloads precisely to reduce the effect of heat. According to the Runtime section of the official internals document, the default pause is 2 seconds in version 6.0 and 5 seconds from 6.1 onward. In its explanation of changes in 6.1, the developer also says the longer pauses reduce thermal throttling and run-to-run variation.
Because the exact version cannot be determined from this comparison chart, it cannot be asserted that every device rested for 5 seconds each time. Even so, the fact that Geekbench 6 is a CPU test that includes pauses is an essential condition for reading the results. It puts load on the device differently from games where heat keeps building, or from tests that run long workloads without a break.
In other words, the closeness of room-temperature and cryogenic scores supports the view that ordinary cooling does not greatly hold back performance in this test. It does not mean the chassis reaches liquid-nitrogen temperatures, nor does it prove that peak performance is maintained for long periods. Delivering enough short-burst CPU performance and being able to keep using that performance are separate things to measure.
If you are choosing an A20 Pro device for gaming or on-device AI, it is useful to look at changes in processing time together with power consumption, using the same room temperature and app settings. It is worth checking the smaller iPhone 18 Pro and the larger Pro Max separately as well. If high performance can be obtained repeatedly under near-everyday conditions, and the drop during sustained processing is small, that speed can be valued as something users can actually use, rather than a record set by cooling the chip to an extreme.