我们采访了苹果芯片背后的男人——Johny Srouji！

Summary

在这段访谈中，苹果公司的Johnny Park讨论了iPhone 16和16 Pro系列的A18和A18 Pro芯片如何支持苹果智能技术。他强调了苹果智能技术对芯片设计的影响，以及8GB RAM在提升用户体验中的重要性。此外，他还提到了苹果在CPU核心数量、微架构设计和热管理方面的策略，以及苹果芯片在提升iPhone摄影能力方面的作用。

Takeaways

• 📱 iPhone 16和iPhone 16 Pro系列是首批搭载苹果智能技术的iPhone，A18和A18 Pro芯片在运行苹果智能技术方面发挥了重要作用。
• 🚀 A18和A18 Pro芯片在设计时就考虑了苹果智能技术，通过增强神经引擎、增加系统内存带宽和容量来提升整体性能。
• 💾 去年iPhone 15 Pro升级到8GB RAM，iPhone 16系列也继续这一配置，这不仅是为了支持苹果智能技术，也为了满足如高端游戏等其他应用的需求。
• 🎮 苹果智能技术是推动iPhone升级到8GB RAM的一个重要因素，但8GB RAM也将极大地促进其他应用的性能，包括高端游戏。
• 🤖 苹果的设计理念是构建统一的CPU架构，无论是手机、iPad还是Mac，以便于软件开发者能够针对单一架构进行开发。
• 🏋️‍♂️ 苹果的CPU设计注重单线程性能和能效，而不是简单地增加核心数量来追求峰值性能。
• 🔧 在决定CPU微架构时，苹果会根据产品需求、电池大小、电源系统和热环境等因素进行综合考虑。
• 🎮 苹果的GPU架构在iPhone和Mac之间保持一致，但会根据不同产品的特点进行实施上的差异化，以保持能效。
• 📹 iPhone 16系列在热管理方面进行了改进，以支持更强大的性能，同时保持了产品的轻薄设计。
• 🎥 苹果的图像信号处理器（ISP）和视频编码器的紧密集成，使得iPhone能够提供Dolby Vision 4K 120fps的视频拍摄能力。

Q & A

苹果A18和A18 Pro芯片如何支持运行苹果智能技术？

A18和A18 Pro芯片通过持续改进神经引擎，增加系统内存带宽，并在整体性能上进行优化，从而支持苹果智能技术。这些芯片从设计之初就考虑了苹果智能技术的需求。

苹果智能技术对苹果硅芯片设计有何影响？

苹果智能技术是苹果硅芯片设计中的一个重要考虑因素，特别是在决定内存容量和带宽时。苹果智能技术是推动iPhone 16 Pro系列升级到8GB RAM的关键因素之一。

为什么iPhone 16系列继续使用2个性能核心和4个效率核心的CPU配置？

苹果的设计理念是提供最佳的用户体验，而不是追求特定的峰值性能。苹果的CPU在单线程性能和能效方面都是行业领先的，因此2个性能核心和4个效率核心的配置足以满足iPhone的需求。

苹果M4芯片相比A18系列在核心数量上有所增加，这是基于什么考虑？

苹果M4芯片用于iPad等产品，这些设备有更大的散热空间和不同的电源供应，因此可以支持更高的频率和性能。苹果通过模拟和性能建模工具来决定不同产品的最佳核心配置。

苹果在CPU微架构设计上是如何做出决策的？

苹果的决策基于提供最佳用户体验的原则，通过大量的数据和性能建模工具来确定最佳的CPU配置。苹果不会为了某个基准测试而设计产品，而是专注于整体的能效和性能。

苹果如何在iPhone上实现高端游戏的性能优化？

苹果通过提供相似的GPU架构在不同设备上实现高性能，同时在实施层面进行调整，以适应不同的电源供应和散热限制。苹果的GPU设计注重能效，以确保在移动设备上也能提供持续的高性能。

苹果在iPhone 16系列中对散热管理做了哪些改进？

苹果在iPhone 16系列中进行了多项散热改进，以提供更好的产品体验。这些改进包括使用新材料和优化组件布局，同时保持设备的轻薄设计。

苹果硅芯片在iPhone的视频拍摄能力中扮演了什么角色？

苹果硅芯片中的图像信号处理器（ISP）和视频编码器在视频拍摄中发挥了关键作用。ISP负责处理图像的色彩和色调，而视频编码器则以高速度压缩视频，使得iPhone能够支持杜比视界4K 120帧每秒的视频录制。

苹果如何在不同设备上保持GPU架构的一致性，同时又满足各自的特点？

苹果通过在不同设备上使用相似的GPU架构，同时在实施层面进行调整，如核心数量和频率，来保持一致性。这样的设计使得开发者可以更容易地将游戏和应用移植到不同的苹果设备上。

苹果如何决定在iPhone中使用特定的散热解决方案？

苹果在决定散热解决方案时，会考虑多种因素，包括材料、设备厚度、电池空间和其他组件的布局。苹果的目标是提供最佳的产品体验，而不是单纯追求某个方面的性能。

Outlines

00:00

📱 iPhone 16系列的A18芯片和Apple智能

在这段对话中，讨论了iPhone 16和iPhone 16 Pro系列的A18和A18 Pro芯片如何支持Apple智能技术。提到了Apple在2017年首次引入神经引擎，并逐年改进以提高性能和能效。A18和A18 Pro芯片进一步增强了神经引擎，增加了系统内存带宽，并改善了整体计算、内存子系统。此外，提到了iPhone 15 Pro去年升级到8GB RAM，而iPhone 16系列也采用了8GB RAM，这不仅是为了Apple智能，也为了支持高端游戏等其他应用。

05:01

🔋 Apple芯片设计的核心原则与性能

这部分内容讨论了Apple芯片设计的核心原则，包括跨不同设备的CPU架构一致性、不追求特定峰值性能的基准测试，以及追求最高能效的性能。强调了Apple芯片在单线程性能和能效方面领先行业，并解释了为何在iPhone上使用2个性能核心和4个效率核心的配置。同时，提到了Apple M4芯片在iPad上增加核心数的原因，以及如何在不同的设备上实现相同架构的不同实现，以适应不同的热环境和电源需求。

10:03

🎮 将AAA游戏移植到iPhone的挑战与解决方案

在这段对话中，讨论了将AAA游戏移植到iPhone时面临的工程挑战，包括移动GPU与桌面或游戏机GPU之间的固有差异，如内存子系统带宽较低、渲染管线不同等。Apple通过保持GPU架构的一致性来简化开发过程，同时在不同设备上进行特定的实现调整。此外，还讨论了iPhone 16系列在热管理方面的改进，包括与产品团队合作，优化整体产品以提供最佳用户体验，而不是单纯追求峰值性能。

15:04

🎥 iPhone在视频拍摄方面的创新与Apple硅片的作用

这部分内容强调了iPhone在视频拍摄能力方面的优势，特别是A18系列芯片在iPhone 16和16 Pro中的作用。Apple在图像信号处理器（ISP）和媒体引擎方面的投资，使得每帧图像都能进行高质量的色彩和色调分析。通过与神经引擎的紧密集成，iPhone能够快速处理图像并提供详细的场景信息。此外，视频编码器能够以高速度压缩视频，使得iPhone能够支持Dolby Vision 4K 120fps的视频录制，这得益于Apple硅片、ISP和软件之间的紧密集成。

Keywords

💡Apple 硅片

Apple 硅片指的是苹果公司设计的自家处理器，如A系列芯片。这些芯片是iPhone和其他苹果设备的核心，负责处理设备的所有计算任务。在视频中，讨论了如何通过改进神经引擎和增加系统内存带宽来优化Apple 硅片，以支持Apple智能和提高性能。

💡Apple 智能

Apple 智能可能指的是苹果公司在设备上实现的人工智能功能，如Siri或机器学习算法。这些智能功能可以提高用户体验，例如通过自然语言处理或图像识别来增强设备的功能。视频中提到，Apple 智能在设计Apple 硅片时被考虑在内，以确保硬件能够支持这些智能功能。

💡神经引擎

神经引擎是Apple 硅片中的一个组件，专门用于处理机器学习任务，如图像和语音识别。视频中提到，苹果公司在A系列芯片中不断改进神经引擎，以提高性能和能效，这对于运行Apple 智能等AI功能至关重要。

💡Transformer 模型

Transformer 模型是一种深度学习模型，广泛应用于自然语言处理任务中。视频中提到，苹果公司在A系列芯片中增加了对Transformer 模型的支持，这表明苹果正在增强其硬件以支持更复杂的AI算法。

💡内存带宽

内存带宽是指数据在处理器和内存之间传输的速度。在视频中，提到了增加系统内存带宽是提高Apple 硅片性能的关键因素之一，特别是在处理大量数据时，如运行Apple 智能或高端游戏。

💡能效

能效是指设备在执行任务时消耗能量的效率。视频中强调了苹果公司在设计硅片时非常重视能效，因为这直接影响到设备的电池寿命和热管理。高能效的硅片可以在不牺牲性能的情况下，提供更长的电池使用时间和更好的散热。

💡单线程性能

单线程性能是指处理器在处理单个任务时的性能。视频中提到，苹果公司专注于提高其硅片的单线程性能，而不是简单地增加核心数量。这样做可以在保持能效的同时，提供出色的性能。

💡图形处理器（GPU）

GPU是负责处理图形和视频渲染的硬件组件。视频中讨论了苹果公司如何通过改进GPU架构来支持高端游戏和视频处理，同时保持能效。这表明苹果公司在设计GPU时，考虑到了移动设备的特殊需求。

💡热管理

热管理是指设备在运行时如何有效地散热。视频中提到了苹果公司在iPhone 16系列中对热管理进行了改进，以确保设备在高负载下也能保持稳定运行。这涉及到选择合适的材料和设计，以优化散热性能。

💡图像信号处理器（ISP）

ISP是负责处理从相机传感器接收到的图像数据的硬件组件。视频中提到，苹果公司在其硅片中集成了先进的ISP，以提供高质量的图像处理，包括色彩、色调和动态范围的优化。这使得iPhone能够提供卓越的摄影能力。

Highlights

苹果在2017年首次引入神经引擎，此后每年都在提升性能和能效。

iPhone 16和16 Pro系列的A18和A18 Pro芯片在设计时就考虑了苹果智能，提升了神经引擎性能。

A18和A18 Pro芯片增加了系统内存带宽，以支持苹果智能和其他高性能需求。

iPhone 16系列继续采用8GB RAM，这一决策不仅由苹果智能驱动，也考虑了游戏等其他应用场景。

苹果的软件团队会优化内存使用，避免浪费，确保最佳用户体验。

苹果M4芯片增加了核心数量，而A18系列则保持了两个性能核心和四个效率核心的配置。

苹果的CPU架构设计旨在提供最佳性能和能效，而不是单纯追求峰值性能。

苹果的产品设计哲学是为特定产品需求定制优化，而不是追求最大核心数量。

苹果的微架构决策基于性能模型和实际数据，以确保最佳用户体验。

苹果的GPU架构在iPhone和Mac之间保持一致，便于开发者进行游戏移植。

iPhone 16系列的热管理系统进行了改进，以提供更好的性能和能效。

苹果在iPhone 16系列中使用了铝板进行热管理，以优化产品的整体设计。

苹果的图像信号处理器（ISP）和媒体引擎紧密集成，提供卓越的视频拍摄能力。

iPhone 16系列能够以每秒120帧的速度拍摄Dolby Vision 4K视频，得益于苹果芯片的强大性能。

Transcripts

Park

Johnny hi Johnny I'm R from J and nice

to have you here today my pleasure good

afternoon my pleasure to meet you so I

heard you are kind of straightforward

guide on Tech topics so maybe we can

just step into the questions ASAP go

ahead please so my first questions being

obviously iPhone 16 and the iPhone 16

Pro Series are probably the first

iPhones after you launch the Apple

intelligence so how do a18 and a18 pro

chips contribute to running Apple

intelligence and for wider pers

perspective has Apple intelligence make

any significant impact on your decision

making when you're designing Apple

silons and maybe overall Hardware yeah U

so you're referring to Apple

intelligence running on app silicon for

the iPhone but we actually introduce and

shipped our first neural engine

implementation in 2017 and since then

we've been improving our neural engine

year-over-year and adding more

performance with power efficiency and

our colleagues from the software team

has been leveraging and utilizing that

engine then when you look at Apple

intelligence running an apple silicon it

utilizes the whole s so but it heavily

also utilizes the neural engine which we

also added support for Transformer

models many years back and that's one of

the reasons that7 Pro can support apple

intelligence so with A8 and 18 Pro we

took it even one step further we kept

improving the neural engine even more

than a17 Pro we added more bandwidth to

the system memory and we added more

improvements across the whole so to to

build a balanced compute memory

subsystem including the capacity and the

bandwidth on the computer so we built it

from the ground up with apple

intelligence in mind you actually

upgraded to 8 gig of RAM on iPhone 15

Pro last year if I'm not mistaken the

iPhone 16 series are also you know

having the upgrade for 8 gigs of RAM

which for me seems like a you know

critical turning point so could you

explain the necessity of you know

bringing more Rams to iPhone is this

decision solely driven by Apple

intelligence or were other use cases

such as gaming also considered so again

our goal is to build the best products

delivering the absolute best user

experience as it relates to Apple

intelligence dram is one aspect and when

we look at what we're building whether

it's silicon Hardware or software we

don't want to be wasteful in many ways

we have lots of data that tells us what

is going to enable a certain feature and

apple intelligence is one of those very

very important features that we want to

enable and we look at different

configurations both for computation and

memory bandwidth and memory capacity and

then we made the right tradeoff and

balance of what actually makes the most

sense so Apple intelligence was a major

feature that led us to believe that we

we need to get to 8 gab but having said

that the 8 gab is going to help

immensely in many other applications

including gaming a high highend gaming

AAA title games and highend gaming on

device so I think it's going to be

really really beneficial the other thing

to keep in mind uh this is one of the

benefits of having the software and the

silicon and the product fully integrated

is that the software team our excellent

software team will optimize not only for

compute they'll also optimize for the

memory footprint of each application so

they don't end up also wasting memory so

we look at all these trade-offs and we

end up with here's what makes sense and

8 GB was the most perfect choice for us

I think would be a par question you know

some of our competitors are having

started to put more core counts into

their CPU designs even the latest Apple

M4 chip has also increased the number of

core counts compared to before and in

contr the a18 series continue to use

let's say two performance cores and four

efficiency cores it's a strategy that

you have been used for quite some times

why did the apple M4 cons considered

increasing the numbers of core counts

while a18 Pro didn't great question so

maybe I will start with our philosophy

and then I'll get to your to answer

specifically your question our

philosophy I I'll give you some uh some

principles one of the principles is we

want to build the same CPU architecture

whether it's for a chip that goes into

the phone or the iPad or the Mac or

other configurations so it's a scalable

architecture same applies to the

graphics neural engine and others now

one of the main benefits for that is for

the software and developers you have a

single architecture that you develop for

the iPhone or the iPad or the Mac so

that's big and by the way another side

benefit is from a team efficiency you

get to design one architecture so that's

one principle some architecture across

different chips the second principle is

given we're not a merchant vendor we

don't really need to Target a specific

Peak Performance for a specific Corner

case Benchmark that you may not even

experience or hit as a customer just in

order to win some Benchmark we care

about again delivering the absolute best

user experience and for that we look at

lots of lots of data of how the devices

the software is using the silicon and

what makes the absolute best use and

therefore we make that R based on that

aspect a third principle which is

important is you want to deliver the

absolute best performance whether it's a

CPU or Graphics but let's talk about CPU

since that's your question with the

highest performance at the best Absolute

Energy Efficiency Energy Efficiency

extremely extremely important for us so

the best single thread and that's key

because what others might be doing other

vendors what they might be doing is they

add multiple cores more and more cores

in order to compensate for not so good

single threat performance so one way to

compensate is you add more cores uh and

therefore you can achieve a certain Peak

Performance at higher power larger die

which means also larger cost so we don't

do that either so now let me answer your

question based on these principles when

you look at the single thread of

performance core across all of our

silicon is the absolute best in the

industry we're leading the industry if

you look at the efficiency cores same

we're at the absolute best we're like

leading in a big way and then when we

look at the configuration whether it's a

selic that goes into the phone or the

iPad or the mac and we have lots of

simulation uh and performance modeling

tools and we look at actual data

and then we take into account for

example the battery size for a product

the power delivery system for a product

the thermal envelope for the product

because overbuilding again is wasteful

for example for the phone we came to the

conclusion that 2 p4e so two performance

course for efficiency course meets the

needs of what that device requires

because we have the absolute best single

thread and the efficiency cause is so

good for other tasks and that

configuration works then your next

question was why did you make different

choice for M4 that goes into the iPad

those have larger thermal envelopes

different Power delivery and therefore

you can push the frequency and the

performance to a different point so

going to my first principle of using the

same architecture across different chips

we make the differentiation in

implementation of frequency points

operating points we enable Peak

frequency in order to enable what we

call best performance to give you the

absolute best performance but we also

look into Energy Efficiency and where

the sustained needs to be and what is

the shape of the curve when you look at

Power performance so that your high

performance maximum Energy

Efficiency okay so the next question is

about the microarchitecture you actually

have a long period of time that micro

architecture of a series CPUs was

steadily updated before A6 I would say

starting with A7 Pro you actually

started to widen the your CPU of course

uh we saw a 9 wide decode on a17 Pro and

depends on the CPU upgrade we saw on

Apple M4 I would say we expect a18

series CPU to widen further so my

question is when you are making those

decisions on CPU micro architectures you

know how PPA trade-offs are made So my

answer actually follows the same

principle that just cover which is you

want to design the best CPU for a

certain silicon that goes into a certain

product and we have lots of data and

performance modeling to tell us what

configuration makes the best use terms

of highest power efficiency and

performance obviously we're not going to

get into micro architecture details on

what plans we have in mind but you can

imagine we have a deep line of CPU micro

architecture not for this year next year

for many years to come and we have

modeling tools including what cash sizes

you need for each of these course per

implementation based on that we make

those decisions and based on what's

going to deliver the best user

experience again not for a certain

Benchmark now in my experience it proved

to be the case that once you do that you

actually end up winning lots of

benchmarks and winning those while

keeping your Energy Efficiency so it

becomes a side benefit but again the

strength that we have being an

integrated part of Apple the full

integration of the software the product

team and the Silicon team where you

design absolutely for what our product

needs not for everyone else gives us the

freedom to optimize what you call PPA

power performance area for the absolute

best Energy Efficiency and area it's

science but it's science combined with

art and you make the proper judgment

codes based on tools and modeling of

what actually going to deliver the best

user experience so it's both okay so

next question is about uh gaming on

iPhones we saw developers bringing more

and more AAA games to iPhone but

obviously there are many inherent

difference between mobile gpus and

desktop or console gpus I say typically

mobile memory subst system would be you

know have much lower bandwidth mobile

GPU rendering pipeline actually differs

for example Apple's GPU are actually

utilizing tile based defer rendering

techniques instead of immediate

rendering and also software development

pipelines are quite you know distinct so

I'm wondering what engineering

challenges did you guys face when

deciding to Port these triaa games to

iPhone and how were they addressed and

by the way again following one of the

principle I just covered which is we

want to build the same GPU architecture

for all of our chips for example you can

see that the GPU architecture

fundamental that we built for the mac

and for the iPhone are very similar the

implementation are different and we can

get to that later but the architecture

is is similar and what that gives again

to developers is they get to port or

Implement a game for the iPhone and the

mac and it's the same porting so that's

great and the fact that our GPU is so

performing that means that for example

when you look at the phone including 18

Pro is that it has longevity meaning you

can looking forward for future games it

will be able to support those at really

high performance low power then when you

look at different implementations

whether it's a phone or a Mac this is

where we make implementation differences

and we're smart about how we do that

during the implementation while keeping

the architecture similar and it's not

only frequency and operating points it's

below that even meaning a lower level of

details but in terms of the functional

archit ecture is is the same of course

you're working with different Power

delivery and thermal envelopes which is

why you can see that appid silicon

shipping on a phone has different or

less GPU cores than appid silicon

shipping in a MacBook Pro but the

fundamental what we call fstp the

shaders Etc are very very similar and

then we operate at different points

again to maintain Energy Efficiency but

the same grade Graphics you get all

across we just talk about the running

games on iPhone and also seral

management is uh crucial factor for that

this year you emphasized improvements on

thermal management so can you

specifically talk about what changes

you've made for the Thal management the

iPhone 16 and the 16 Pro Series and

could you talk through decision making

and also how does it differs from

previous thermos systems again we build

products and the idea is to deliver the

best products and it's not only about

theque performance for graphics at

certain termal envelope and again this

is where my team and the product team

work hand in hand ahead of time about

what is the absolute best decision we

can make at the product level of course

in terms of silicon you want more

thermos you want bigger battery but that

has to off on other aspects so we need

to manage that piece and we highlight in

the event this morning about some of the

many thermal improvements that were were

made into the iPhone 16 series and that

obviously benefits theid silicon that is

enabling those iPhones but then you go

back to the Energy Efficiency the fact

that we have the most energy efficient

chips it helps us and again when you

don't chase Peak Performance at higher

power and not so energy efficient versus

sustained and best in performance you

get to optimize for the overall system

obviously we can get bit deeper into

that you guys already implemented if I'm

not wrong aluminum plate inside phone

actually in the industry a lot of people

are using let's say Vapor Chambers and

other Solutions so when you are deciding

new thermal solution what kind of as a

mindset did you go through what kind of

factors do you considers you know again

the mindset is who want to deliver the

best product not necessarily one thing

the best product optimized for the

absolute best user experience for

example you can deliver a better thermal

envelope if you have thicker devices you

use different

materials the example you gave is one

more example but that's one tradeoff we

believe if you take it to the extreme

doesn't actually benefit the product

because it has implications on the form

factor and the ID so we take all

consideration to account including the

materials and how thick you can allow

that product we want it thinner how much

space you give battery and other

components even the placement of

components on the MLB make a difference

so it's very thoughtful thorough process

that takes into consideration many many

aspects but the absolute number one

priority is building for the absolute

products not just thermal so that I can

deliver a Peak Performance and saying I

win this Benchmark we want to deliver

the best absolute product and and the

Energy Efficiency Focus we have all

across enables all of this and again

it's an optimization at the product

level okay so the last question let's

talk about video shooting on iPhone now

so uh obviously iPhone have very great

humaning capability compared to

competitors what kind of roles did Apple

silicons made in that kind of process

well maybe for example A8 series inside

iPhone 16 and 16 pros of course this

started many many many years back uh

we've been investing in our image

silicon processor which is the ISP and

same applies for other media engines so

we've been investing in that space we

believe we have the best absolute custom

silicon that is built for these devices

for example if you look at each frame

that passes through our camera ISP it

gets analyzed for things like color and

tone when you look at the image signal

processing as a pipeline it starts with

a an excellent image information that

gets into the high speed and then it

gets defined even further by dedicated

streaming machine learning based

processing that have been trained on

millions of photos before that and those

run on the new engine so you can see the

tight integration between the new engine

and the IP in this case now the new

engine also provides detailed semantic

information about the overall scene

including the subjects various subjects

and the intent of the photographers now

what this does it enables the camera to

deliver a wider range of dynamic

information based on the scene and the

materials in the scene Etc and we get to

do that very very fast at a high rate

124k frames per second so that's great

then post ISP as as that gets processed

we have another media engine called the

video encoder that compresses the video

at 1 billion pixels per second so when

you combine both you get to a very high

rate that's how we enable Dolby Vision

4K 120 frames per second sometimes you

can see that others you know our

competitors can generate through lots of

software processing a good frame a good

picture at one time but we're first in

terms of delivering a video for D Vision

Video 4K 120 capture um and that's

thanks to the tight integration of the

Silicon different IPS on the software

and the camera control it's like the

physics camera control and the Silicon

including the ISP they all work together

I think that's all the question we have

today so thank you my pleasure thank you

thank you thank