Check the full review for a detailed model comparison table
"The third part of the processor triumvirate – and likely where the bulk of Apple’s transistor budget has gone – is the neural engine, Apple’s name for their block responsible for processing (inferencing) neural networks for AI. Relative to the A11, Apple is promoting some huge increases in both the basic block size and in the resulting performance. The latest neural engine is described as an 8-core design, up from 2 cores for the A11. In turn, peak performance has jumped from 600 GigaOPS to 5 TeraOPS, an increase of over 8x.Apple Announces the 2018 iPhones: iPhone XS, iPhone XS Max, & iPhone XR -- AnandTech
Given the core count, I have no doubt that part of this outright comes down to Apple laying down a lot more silicon for the task. Which goes to show just how important Apple considers the neural engine and the workloads that run on it towards current and future iPhone applications. However the latest version of the neural engine also adds what Apple calls “multiprecision” support, better known as variable precision. Smaller data formats lose precision, but they can be processed more efficiently. So Apple’s 5 TOPS claim is almost certainly based on this, which is an important distinction because not all workloads can be run at a lower precision. So the real-world performance difference may not be as great as it looks on paper."
No comments:
Post a Comment