A20 Pro’s Neural Engine can outrun its own GPU — on the right work
Max Weinbach’s Friday note: peak compute from the dual 16-core ANE on A20 Pro and M6 beats the GPU when the job maps. Prefill yes. Pretty frames no.
Short1 minEasy
Eli Navarro1 min read
- Apple
- A20
- M6
- Chips

Photo: WccftechSource page
The morning chip story was the node: 2nm, iPhone 18 Pro, a foldable to spend it on. The Friday evening add is a narrower boast. Max Weinbach wrote that peak throughput from the dual 16-core Neural Engine on A20 Pro and M6 is higher than the GPU on the same silicon — if the workload maps to the ANE. Wccftech amplified it. Smaller models around 3 billion parameters and prompt prefill on quantized LLMs are the examples. Games and animation are not. Send those to the 7-core GPU or they will crawl.
Single-token generation can lose the ANE’s advantage to dispatch overhead, which is why CPU or GPU still win some of the chat loop. The useful reader takeaway is not a benchmark screenshot. It is that Apple’s first 2nm phone is being sold as an on-device model machine, and the part of the chip that does that work is no longer the side salad. Phones are not out yet. The claim will get a real test when the first 18 Pro units hit a prompt that is not a keynote reel.


