Apple A20 Pro Performance: How Fast Is the 2nm iPhone Chip?

The A20 Pro posts 24–33% higher CPU scores than the A19 Pro, while Apple claims up to 40% faster graphics and sustained performance. A local LLM ran more than twice as fast in one on-device test, but image generation showed a much smaller generational gap.

The A20 Pro does not have one universal performance number

Running the same local AI model on an iPhone 17 Pro and iPhone 18 Pro produced the headline result: the newer phone generated its answer more than twice as fast. Because the model ran entirely on each phone, server latency was not responsible for the difference.

That does not mean every A20 Pro workload is twice as fast. CPU benchmarks show a large but smaller generational gain. Tom's Guide measured 4,735 single-core and 12,800 multi-core in Geekbench 6 on the iPhone 18 Pro, up about 24% and 28% from the iPhone 17 Pro. Geekbench 7 scores rose by about 28% and 33%.

The Geekbench 7 multi-core result lands in the same broad range as a base M3 MacBook Pro from 2023. The comparison mixes a publication's iPhone test with aggregated Mac results, so it does not make the two devices equivalent. It shows how far phone CPU performance has moved within the same benchmark.

Why the local LLM gain was larger than the CPU gain

On-device AI is not a CPU-only workload. The CPU schedules work and handles general logic, the GPU processes large groups of similar calculations, and the Neural Engine handles supported machine-learning operations.

The A20 Pro expands all three paths. Its six-core CPU and seven-core GPU include Neural Accelerators, while the Neural Engine moves to a Dual 16-core design with 32 cores in total. Apple rates the new Neural Engine at twice the on-device AI compute of the A19 Pro.

Putting accelerators inside the CPU and GPU reduces the need to hand every small AI operation to a separate block and wait for the result to return. Large parallel operations can still go to the GPU or Neural Engine, while smaller compatible operations stay closer to the part of the chip already handling the surrounding work.

This distributed design helps explain why a compatible local LLM can improve by more than the CPU benchmark alone suggests. The result depends on how the runtime divides the model across the available hardware.

A 50% wider memory path matters during token generation

Apple says the A20 Pro has 50% more memory bandwidth than the A19 Pro. This is not a 50% increase in storage or RAM capacity. It means the chip can move more data between memory and its compute blocks in the same amount of time.

Local language models repeatedly read large sets of model weights from memory. Fast compute units still stall when the required data arrives too slowly. More bandwidth reduces that wait.

Prompt processing and token generation also stress the system differently. The initial prompt can be processed with substantial parallelism. Once generation starts, the model selects the next token and repeatedly streams weights from memory. At that stage, bandwidth can limit output speed even when raw compute is available.

The more-than-2× result in the local LLM test is therefore not a Neural Engine result in isolation. It reflects the combination of wider memory bandwidth, CPU and GPU accelerators, the Dual Neural Engine and a runtime capable of placing work on them effectively.

Local image generation did not produce the same 2× leap

The same test also ran an image model locally, creating and saving more than 20 images in sequence. Producing 20 images took about 11 minutes. The iPhone 18 Pro did not pull away from the previous generation by as much as it did during language-model output.

Both tasks are labelled AI, but their bottlenecks differ. Autoregressive text generation often streams model weights repeatedly for each token and can benefit heavily from additional bandwidth. Image-generation performance can depend more on GPU throughput, model architecture and how the app's runtime schedules each stage.

The result is a useful warning against treating TOPS, Neural Engine core counts or a single AI demo as a universal performance rating. Different models can produce very different gains on the same chip.

Apps must be updated to use the new acceleration paths

Hardware accelerators do not make every existing app faster by the same percentage. The app decides which operations run on the CPU, GPU or Neural Engine, while iOS and the model runtime must support those paths.

Software tuned for the A20 Pro may improve after launch as developers adopt its accelerators and memory behavior. A poorly matched runtime can leave new hardware underused. A workload that is limited by data movement also needs smarter reuse and model execution, not just faster arithmetic.

Launch measurements show what current apps and runtimes achieve, not a fixed ceiling. The performance users actually see will depend on how quickly apps and frameworks learn to use the new hardware.

The 2nm node works together with new packaging and cooling

The A20 Pro is Apple's first 2nm iPhone chip. TSMC's N2 platform replaces FinFET transistors with nanosheet transistors, allowing the gate to control the channel from more sides. TSMC's platform target is 10–15% more speed at the same power or 25–30% lower power at the same speed than N3E, with more than 15% additional chip density.

Those are process-platform targets, not direct A20 Pro product measurements. Apple's CPU, GPU, Neural Engine and memory architecture determine how that transistor budget is used.

Packaging and cooling determine how long the chip can hold its speed. Apple places the silicon die beside the memory, moving the memory out of the chip's thermal path and allowing the A20 Pro to connect directly to a next-generation vapor chamber. The layout is inspired by Apple's M-series packaging.

The iPhone 18 Pro vapor chamber has three times the surface area of the previous design. Apple claims the new materials and cooling system enable up to 40% better sustained performance than the iPhone 17 Pro. That gain matters most after the phone has been running a demanding game or local AI model long enough to heat up.

The A20 Pro is designed for sustained on-device workloads

The A20 Pro is substantially faster than the A19 Pro, but the size of the gain depends on the workload. CPU scores rise by roughly 24–33%. Apple rates graphics and sustained performance at up to 40%. Its on-device AI compute rating doubles, and one compatible local LLM generated text more than twice as fast.

Image generation did not show the same jump. The chip does not make every AI task twice as fast. Memory-heavy language models, accelerator-aware apps and long-running workloads are more likely to expose its largest advantages.

The A20 Pro's most important change is not the 2nm label by itself. It is a system designed to keep on-device AI running at high speed. More compute, 50% more memory bandwidth and a larger cooling path work together. Local LLMs, demanding games and sustained creative workloads can show larger gains when their apps use the new acceleration paths.

Frequently asked questions