Apple Wiki
Register
Advertisement
Apple Wiki
Apple M1 simplified schematic

The 16-core Neural Engine in the schematic of the Apple M1.

Apple's Neural Engine (ANE) is the marketing name for a group of specialized cores functioning as a neural processing unit (NPU) dedicated to the acceleration of artificial intelligence operations and machine learning tasks.[1] They are part of system-on-a-chip (SoC) designs specified by Apple and fabricated by TSMC.[2]

History[]

Core ML icon

Icon of the Core ML API

The first Neural Engine was introduced in September 2017 as part of the Apple A11 "Bionic" chip. It consisted of two cores that could perform up to 600 billion operations per second for real-time processing of machine learning algorithms dedicated to features such as Animoji and Face ID. However, an API was not yet available for 3rd-party developers. The Neural Engine in the Apple A12 (also called "Bionic") was expanded to eight cores that could process up to 5 trillion operations per second. An API named Core ML was released in 2018 to allow developers to take advantage of the Neural Engine, which ran up to nine times faster on the A12 while using one-tenth of the power of the version in the A11. The 8-core Neural Engine within the Apple A13 that was released in September 2019 was 20% faster while using 15% less power.[3]

The Apple A14 was released in October 2020 with the number of Neural Engine cores doubled to 16, providing the ability to perform up to 11 trillion operations per second.[4] The Apple M1 was released in November 2020 with improved overall performance and power efficiency, but a Neural Engine that was largely unchanged from the A14.[5] The Apple A15 was announced in September 2021 with 16 Neural Engine cores like earlier models, but with the ability to perform up to 15.8 trillion operations per second.[6] However, Neural Engine performance in the Apple M1 Pro and Max, introduced in the following month, was unchanged from the lower performance of the original M1.[7] The Apple M1 Ultra was introduced in March 2022 which used a previously unrevealed interconnect called UltraFusion to effectively double the number of cores and performance over that of the M1 Max.[8]

In June 2022, Apple introduced the Apple M2 with 16 Neural Engine cores that could deliver over 40% faster performance than the Apple M1 with the same number of cores, in line with that of the Apple A15.[9] In September 2022, Apple introduced the Apple A16 with 16 Neural Engine cores on a new 4 nm process that could handle nearly 17 trillion operations per second.[10] The Apple M2 Ultra was introduced in June 2023, using UltraFusion to double the number of cores and performance over that of preceding M2 chips.[11]

In September 2023, Apple introduced the Apple A17 Pro with 16 Neural Engine cores on a new 3 nm process that could handle 35 trillion operations per second.[12] The Apple S9 was also introduced with 4 cores to support new watchOS features.[13] In late October 2023, the Apple M3 was introduced on the same process, but with 16 cores that could handle 18 trillion operations per second.[14]

Neural Engine version history[]

SoC Introduced Process Neural cores Peak ops/sec. Note
Apple A11 Sep. 2017 10 nm 2 600 billion First implementation.
Apple A12 Sep. 2018   7 nm 8 5 trillion 90% lower power consumption.
Apple A13 Sep. 2019   7 nm* 8 6 trillion 15% lower power. Enhanced 7 nm (N7P).
Apple A14 Oct. 2020   5 nm 16 11 trillion
Apple M1 Nov. 2020   5 nm 16 11 trillion Comparable to Apple A14
Apple A15 Sep. 2021   5 nm* 16 15.8 trillion Enhanced 5 nm process (N5P).
Apple M1 Pro
Apple M1 Max
Oct. 2021   5 nm 16 11 trillion Same as original Apple M1.
Apple M1 Ultra Mar. 2022   5 nm 32 22 trillion 2x Apple M1 Max.
Apple M2 Jun. 2022   5 nm* 16 15.8 trillion Comparable to Apple A15.
Apple A16 Sep. 2022   4 nm 16 17 trillion Improved power efficiency.
Apple M2 Pro
Apple M2 Max
Jan. 2023   5 nm* 16 15.8 trillion Same as original Apple M2.
Apple M2 Ultra Jun. 2023   5 nm* 32 31.6 trillion 2x Apple M2 Max.
Apple A17 Pro Sep. 2023   3 nm 16 35 trillion First 3 nm process.
Apple S9 Sep. 2023   5 nm* 4 Derived from Apple A15.
Apple M3 Oct. 2023   3 nm 16 18 trillion Derived from Apple A17 Pro.
Apple M3 Pro
Apple M3 Max
Nov. 2023   3 nm 16 18 trillion Same as original Apple M3.

References[]

See also[]

  • Velocity Engine, older vector processing technology used in the PowerPC G4 and G5.

External links[]

Articles[]

IPod Nano 6 This article is a semistub. You can help by expanding it with some more information.
Apple Silicon
A series : A4 | A5 · A5X | A6 · A6X | A7 | A8 · A8X | A9 · A9X | A10 · A10X | A11 | A12 · A12X · A12Z | A13 | A14 | A15 | A16
M series : M1 · M1 Pro · M1 Max · M1 Ultra | M2 · M2 Pro · M2 Max · M2 Ultra
Motion M series : M7 | M8 | M9 | M10 | M11
R series : R1
S series : S1 · S1P | S2 | S3 | S4 | S5 | S6 | S7 | S8
T series : T1 | T2
W / H series : W1 · H1 | W2 | W3
U series : U1
Advertisement