未登录,请登录后再发表信息
最新评论 (0)
播放视频

抢占芯片市场赛道——制造完美的电脑芯片

The Race to Build a Perfect Computer Chip

There are various estimates
关于数字经济
of how big a drain on the world’s resources
对世界资源的消耗程度
the digital economy is.
有多种估计
One data point is that 1% of the world’s carbon emissions
一项数据指出 全世界1%的碳排放
are created by people just streaming video
是人们在手机或个人电脑端
over their phones or on their PCs.
播放流媒体视频造成的
Smartphones, computers, data centers,
据估计 智能手机 电脑 数据中心
and all the digital activities we take for granted
以及所有我们习以为常的数字业务
are estimated to use around 7% of the world’s electricity.
消耗了全世界约7%的电力
And that’s only going to get worse
随着人们对电子设备的依赖增加
as we rely more and more on our devices.
这种情况只会越来越糟
Probably the best way to look at that
了解上述情况的最好方法也许
is what’s going on in data centers.
是了解数据中心里发生了什么
These are these giant factories full of computers
这些巨型工厂里满是计算机
that are processing
正在处理
all of the information that you generate.
人们生成的所有信息
These giant data centers
也许到这个十年结束时
are on a course to be using much more electricity
这些大型数据中心的用电量
by perhaps the end of the decade.
将大幅提升
Clearly, that kind of a drain is just not sustainable.
显然 这样的消耗不可持续
We need new solutions.
我们需要新的解决方案
We need novel solutions.
让人眼前一亮的方案
Scientists and startups around the world
世界各地的科学家和初创公司
are developing low-energy computer chips
正在开发低能耗的电脑芯片
which could be better for the environment
这些芯片对环境更友好
and upgrade the intelligence of our digital devices.
让数字设备更智能
Everything’s tied to computers getting better.
与电脑相关的一切都会越变越好
It could be finance,
像是金融
it could be chemical engineering, name your field.
化学工程或任何你所知的领域
So my goal with since founding the company
自公司成立以来
has been to develop a new computing paradigm,
我的目标一直是开发新的计算模式
a new technology platform,
创建新的技术平台
that will allow computers to continue to make progress.
这些会让电脑本身不断进步
It’s not only about design
这不仅仅关乎设计
but also sending them for fabrication, testing them
还包括生产制造 测试
and making system prototypes for various application.
以及为各种应用制作系统原型
The focus of the application
这些应用的焦点
would be reducing the energy consumption
应当是把能耗
to the minimum level.
降到最低
It’s a very high risk, high reward type of problem.
这是一项高风险 高回报的挑战
So if we’re successful,
如果成功的话
then we’ll transform this industry.
我们将改变这个行业
That’d be this enormous leap in this field.
这将是该领域的巨大飞跃
低能耗芯片
From space hardware to toasters,
从飞船硬件到烤面包机
almost everything in modern life
现代生活中的一切
depends on silicon-based semiconductors called chips.
几乎都依赖名为芯片的硅基半导体
They’re the things that make electronic items smart.
芯片让电子元件变得智能
At a very fundamental level,
在最基础的层面上
semiconductors are made up of what’s called transistors,
半导体由所谓的晶体管组成
and these are the absolutely microscopic digital switches,
晶体管是种极其微小的数字开关
the on and offs.
控制信号的开与关
On or off means zero or one in the digital realm.
在数字领域 开或关代表着0或1
Say a modern graphics chip
比如在一个现代图形芯片上
will have tens of billions of transistors on it
会有数百亿个晶体管
and this is something that’s the size of your thumbnail.
而芯片只有你的大拇指指甲那么大
So how do we get all of these tiny switches
那么我们要怎样把所有微型开关
onto something relatively small?
都装到相对小的芯片上呢?
Well, we do that by building up layers of material
我们是这样做的:在硅盘上
on a disk of silicon,
铺上材料层
scraping off patterns into those materials,
在材料层中刻出电路图
layering on other materials
再铺上其它材料
until we have these tiny circuits
直到完成赋予芯片功能的
that give the chip its function.
微型电路
For half a century,
半个世纪以来
the semiconductor industry has made chips
半导体工业通过缩小晶体管尺寸
exponentially faster and better
使芯片的处理速度和质量
by shrinking the size of transistors,
成倍提升
turning room-sized computers
把一间房那么大的电脑
into pocket-sized smartphones,
变成口袋大小的智能手机
propelling the digital revolution.
推动了数字革命的兴起
Clearly, new technology is presenting
显然 新科技
both management and staff
为管理层和员工
with an ever-growing range of choices
未来如何办公 在哪办公
about how and where work will be done in the future.
提供了更广泛的选择
But traditional computing
而传统计算
is reaching its limit.
已至极限
State-of-the-art mass produced chips
实现规模生产的最先进的芯片
are made with what’s called 5 nanometer technology,
采用了所谓的5纳米技术
a dimension that smaller than a virus.
其尺寸比病毒更小
And materials inside some devices
有些设备内部的材料
are already one atom thick,
厚度只有一个原子的大小
meaning they can’t be made any smaller.
这意味着它们不能再缩小了
And we’re reaching the limits of physics.
我们已达到物理学的极限
We’re reaching the limits of energy density.
触及能量密度的极限
The increasing consensus in the chip industry is
芯片行业日益达成的共识是
that these advances that we’ve got from silicon
由硅带来的行业进步
are beginning to come to an end.
正迈向终结
So that if we really want to take advantage of
因此 若我们真的希望充分利用
the full potential of artificial intelligence
人工智能的潜力
and the absolute ocean of data
和如今我们创造的
that we’re all creating today,
海量数据
we’re going to need new materials,
我们就需要新的材料
we’re going to need new programming models,
需要新的程序模型
we’re going to need new versions of computers.
需要新的电脑版本
一、碳纳米晶体管 威斯康星州 麦迪逊
My personal first time about learning
我个人初次学习
and discovering what a carbon nanotube was
和发现什么是碳纳米管
was when I was just about finishing up college, undergraduate.
是在我即将本科毕业之时
And I saw a presentation
我看了
by the vice president of research of IBM
IBM研发部副总的发布会
and immediately I was,
当时我就想
thought it was very special.
这太特别了
A carbon nanotube is a material
碳纳米管是一种
made entirely of carbon atoms.
完全由碳原子构成的材料
The carbon atoms are arranged in a tube.
碳原子呈管状排列
Imagine taking a sheet of carbon atoms
想象一下 把一大片碳原子
and rolling them into a tube
卷成管状
and then making that tube
然后将管子
smaller and smaller and smaller and smaller
越卷越小 越卷越小
until the diameter of that tube was just a few atoms.
直到管子的直径只有几个原子大小
It’s this extremely small,
碳纳米管的直径非常小
which is about 100,000 times
大约是眼睫毛直径的
smaller than the diameter of an eyelash
十万分之一
and a hundred times smaller than a virus.
是病毒大小的1/1000
The really cool thing about nanotubes is that
纳米管最妙的地方在于
they conduct electricity better than just about any other material
其导电性比至今发现的
that’s ever been discovered.
其它任何材料都好
Electrons will move along the length of a nanotube
电子在纳米管里移动的速度
faster than they do in silicon.
比在硅中更快
And that means you can get ultimately faster switching
这意味着最终 你能在开与关的状态间
between on and off states.
更快速地切换
You can make faster computer chips.
制造出处理速度更快的电脑芯片
You could turn them on and off with less voltage.
用更低的电压实现开与关
And that means they use less electricity,
这代表碳纳米管用电量更少
less power than silicon.
能耗比硅更低
In theory, nanotubes will be able to do
理论上说 纳米管的性能
a thousand times better than silicon.
能够达到硅性能的1000倍
Same computational capabilities,
同样的计算力
1,000 times less power.
硅的能耗却是纳米管能耗的1000倍
Another advantage is that carbon nanotubes
碳纳米管的另一个优点是
can be processed at a low temperature,
可以进行低温加工
so layers of nanotubes can be built on top of one another.
因此 纳米管可以层层堆叠
Silicon has to be processed at an extremely high temperature,
而硅必须在极高的温度下进行加工
which makes 3D layers much harder.
这会使立体分层更加困难
If you really think about a city,
认真想想一座城市
what happens when you run out of real estate in a city
如果建筑用地耗尽之后会发生什么
is you build up, you build skyscrapers,
你会建造摩天大楼
you build into the third dimension.
构建第三维度
And so if you can’t make the transistors smaller,
因此如果无法再缩小晶体管
then you could improve your computer chip
你可以做更多的晶体管
by making more transistors,
在其他晶体管上
by making multiple layers of transistors
制造多层晶体管
on top of other transistors.
来改进电脑芯片
Since the 1990s
自20世纪90年代
when nanotubes were first invented in Japan,
日本率先发明纳米管以来
different methods have been developed to mass produce them.
人们用了各种方法来实现其规模化生产
But every process creates two types of nanotubes,
但每个生产过程都会产出两种纳米管
metallic ones and semiconducting ones.
金属纳米管和半导体纳米管
A metallic nanotube like a copper wire,
金属纳米管就像铜线
it’s stuck in the metallic state.
它稳定在金属状态
You can’t switch it,
无法切换
and that kills a circuit.
这会破坏电路
So we need only semiconducting nanotubes
所以我们只需要半导体纳米管
to really make nanotube electronics work.
来真正实现纳米管电子电路的运行
So molecules and polymers
分子和聚合物
that can be mixed in with the carbon nanotube powder,
可以与碳纳米管粉末混合
which is this tangled mess of nanotubes.
就是这一堆交错杂乱的纳米管
And those molecules and polymers will stick
这些分子和聚合物只会附着在
to just the semiconducting ones
半导体纳米管上
and not the metallic ones,
而不会附着在金属纳米管上
or they’ll stick differently.
或者说它们的附着方式不同
And then you could sort the nanotubes
然后你可以根据这些差异
and separate them based on these differences.
将纳米管分开和分类
Initially, in a powder,
最初在粉末中
there’s about 67% of the nanotubes are semiconducting.
含有大约67%的半导体纳米管
But using these chemical approaches,
但使用上述化学方法
we could extract over 99.99%
我们可以提取纯度99.99%以上的
of the semiconducting nanotubes.
半导体纳米管
After the semiconducting nanotubes have been extracted,
提取出的半导体纳米管
they float around in a solution,
漂浮在溶液中
and so the next challenge is to line them up neatly
因此下一个挑战是将它们整齐地
on a silicon wafer,
排列在硅晶片上
which can then be turned into a computer chip.
再将其制成电脑芯片
Carbon nanotubes really got me hooked and excited
碳纳米管从一开始就让我
right from the beginning.
着迷和兴奋
I mean they really have the opportunity to
我的意思是它们的确有可能
potentially revolutionize electronics.
彻底改变电子产品
This challenge of aligning them
自从发现碳纳米管以来
in these highly packed, highly aligned arrays
将它们排列在这些高度集成
has really been frustrating
高度校准的矩阵中
the field of carbon nanotube electronics
这个难题一直困扰着
since their discovery.
碳纳米管电子研究领域
During my PhD,
读博期间
I was kind of tasked with actually determining,
我研究的方向是确定
you know, how we can better align carbon nanotubes.
如何更好地排列碳纳米管
We just found, you know, kind of by accident
有些偶然地 我们发现
that when we layer carbon nanotube ink or inks
当我们将碳纳米管溶解于
where carbon nanotubes are dissolved in solutions
有机溶剂等溶液中
like organic solvents,
将碳纳米管油墨进行分层时
if we layer that carbon nanotube ink on water,
如果我们将碳纳米管油墨与水分层
then instantaneously those carbon nanotubes
碳纳米管会瞬间
will collect and confine at that ink water interface.
聚集并锁定在油墨与水的分界面上
And that high ordering and induced ordering by themselves
这种极其有序的诱导排列本身
really helps to align the carbon nanotubes.
的确有助于碳纳米管的排列
What you’re seeing here is a aligned array of
这里你看到的是
hundreds of thousands of carbon nanotubes.
数十万个碳纳米管的排列矩阵
As you can see, individual nanotubes,
如你所见 这些浅色部分是
these light-colored regions,
一个一个的纳米管
and then the dark-colored region is the substrate.
深色部分是基底
They’re all really well-aligned
它们和谐共存
with respect to each other.
有序排列
Occasionally, there’s a misaligned nanotube in this,
偶尔 其中会有个别纳米管错位
then we need to improve our manufacturing process
这就需要我们改进制造工艺
to eliminate those instances.
来杜绝这些情况
The biggest current challenge
在用集成碳纳米管代替硅以前
is being able to achieve that high alignment
硅晶片大小的行业标准是300毫米
in extremely uniform arrays
在这么小的硅晶片上
across, you know, 300-millimeter wafers,
极其均匀的矩阵排列中
which is an industry standard of silicon wafers,
实现高度校准
before they start changing silicon to integrate carbon nanotubes instead.
是目前最大的挑战
If we can overcome these challenges,
如果我们能战胜这些挑战
if we can make these aligned nanotube wafers,
制造出高度校准的纳米管晶片
the major players in industry will really jump into this field
行业大拿将真正投身于这一领域
and it could just progress at that point could be very rapid.
该领域的进步也许会十分迅速
The technique isn’t perfect yet,
这项技术还不完美
but it’s already an advance for carbon nanotube research.
但它已经是碳纳米管研究的一大进步
Michael and Katherine have founded a company
迈克尔和凯瑟琳成立了一家公司
with an aim to solve remaining challenges.
旨在解决剩下的问题
But many more breakthroughs are needed
但在纳米晶体管
before nanotube transistors
有机会取代硅之前
have even a chance of replacing silicon.
我们还需要更多的技术突破
It absolutely shows promise,
这肯定有希望实现
but it’s been showing promise for 20 years.
但这种希望已经持续了20年
There are many issues,
仍然存在很多问题
you know, how robust are these devices,
像是这些设备的耐用度
but more importantly can you manufacture them?
但更重要的是 你能生产出来吗?
The semiconductor industry
半导体产业
is based on silicon transistors,
以硅晶体管为基础
and enormous sums have already been invested in infrastructure
我们在生产硅晶体管的基础设施上
to manufacture that technology.
投入了巨额资金
Plants have a price tag of $20 billion,
设备造价高达200亿美元
take years to build
建造耗时数年
and need to be run for 24 hours a day
设备需每天24小时运行
to turn a profit.
才能实现盈利
Change will be difficult without a guarantee
如果不能保证降低碳纳米管造价
that carbon nanotubes will be cheaper.
改变就很难实现
Silicon’s been around a long time.
硅已经存在很长一段时间了
People know how to use it,
人们知道如何应用硅
they know how to program it,
如何用硅编写程序
they know how to mass manufacture it.
如何实现硅的规模化生产
It’s still, in terms of the economics, the winner.
从经济角度来看 硅仍是赢家
And until those economics change,
在经济因素改变之前
nothing is going to replace it.
什么都不能取代硅
二、光子芯片 马萨诸塞州 波士顿
I’ve always been obsessed with computers.
我一直痴迷于电脑
You know, I had an opportunity to work in industry
我曾有机会投身这个行业
at a large semiconductor company for a number of years.
在一家大型半导体公司工作多年
And I got to see there some of the fundamental challenges
在那 我看到了一些与不断缩小晶体管
associated with continuing to shrink transistors
相关的基本问题
and I found that to be not a very exciting future.
我发现 未来不容乐观
So my goal with Lightmatter since founding the company
自Lightmatter公司成立以来
has been to develop a new computing paradigm,
我的目标一直是开发新的计算模式
a new technology platform
创建新的技术平台
that allows you to do calculations using photonic devices.
让你可以用光子设备进行计算
When electronics uses electrons,
电子学中应用电子时
that’s the medium of transfer,
电子就是传输媒介
that’s what represents the data.
它代表着数据
Photonics is using photons,
光子学应用的是光子
which is the basic constituency of light.
光子是光的基本组成部分
So for example, a fiber optic cable
例如 横跨太平洋的光纤电缆
that spans the Pacific Ocean
就是利用光
is using light to transmit information.
来传输信息的
Why is it using light?
为什么用光呢?
Well, it’s because the energy required
这是因为
to shove electrons along a copper cable
用铜电缆传输电子
would be absolutely enormous,
耗能巨大
and you would just have too much signal decay.
并且信号会大大减弱
So if you can convert that information into photons,
如果能将信息转换成光子
you can send it faster and with less energy overhead.
那么信息传送速度会更快 能耗支出会更低
So that’s obviously a desirable thing.
这显然是人们想要的
There’s nothing faster than the speed of light.
没有什么能快过光速
So when you think about the latency
当你想到查询问题
between when you make a query
和光子电脑给出答案之间
and when a photonic computer gives you an answer,
存在延迟时
it’s really the fundamental limit
这的确从根本上
for how fast you could do these things.
限制了你做这些事的速度
Electrical wires,
电线
there’s a certain amount of time that it takes for them to turn on and off.
需要一定时间来打开和关闭
And every time you do,
每次你这样做
it takes a lot of energy to do this,
都需要很多能量
to turn this wire off and on.
来关闭和打开这根电线
Now with a photonic computer and photonics in general,
总的来说 现在有了光子电脑和光子学
you have a type of wire that doesn’t take very long at all,
你就拥有了一种几乎没有延时的导线
femtoseconds maybe, even attoseconds to turn on.
可能只需几飞秒 甚至几阿秒就能打开
It takes almost no energy other than any light that’s lost
除了光信号通过导线传输时
while the optical signal is propagating through that wire.
光本身的损耗 几乎无需消耗任何能量
You have another great property with photonics
光子学的另一个优越性在于
which is that you can send multiple data streams
你可以通过光子导线
through this photonic wire,
同时传输多个
encoded in different colors or wavelengths,
以不同颜色或波长编码的
all at the same time.
数据流
So you have massively parallel computation and communication
你可以通过这些近乎完美的导线
through these nearly perfect wires.
进行大规模并行计算与通信
The idea of silicon photonics
硅光子学的概念
has been around again for a long time.
同样已经出现很久了
The problem comes with changing
问题在于
those electrons-based signals into photons.
如何将这些基于电子的信号转变成光子信号
So the traditional way of doing that
传统的做法
is to have all of these other pieces of equipment
是使用所有昂贵且能耗高的
that are expensive and use lots of energy to do that.
其他设备来完成的
Silicon photonics is,
硅光子学的办法是
well, maybe we can just design a chip
也许我们能设计一种芯片
that directly deals in both electrons and photons
可以直接处理电子和光子
that we can have, you know, the benefits of both worlds.
这样 我们在光子学 电子学两大领域都能获益
And that’s what was supposedly on the cusp of doing.
而这正是我们应该做的
Again, it’s been promised for a long time,
同样的 人们希望实现它很久了
it’s gonna change the world,
一直以来
or has been about to change the world for a long time,
人们承诺这将改变或已然改变世界
and it hasn’t yet.
但仍未实现
But Lightmatter still believes
但Lightmatter公司仍相信
silicon photonics is going to happen,
硅光子学会成为现实
and it’s harnessing existing laser technologies
公司正利用现有的激光技术
used in the telecoms and semiconductor processing industries.
将其应用于电信和半导体加工业
Their chips are specifically being made for AI applications
其芯片专为聊天机器人和自动驾驶汽车等
like chat bots and self-driving cars.
人工智能应用程序而设计制造
The startup plans to launch their first photonic computer
这家初创公司计划在今年晚些时候推出
later this year.
其第一台光子电脑
One of their current products is an interconnect,
他们目前的一种产品是互连装置
which enables chips to talk to each other.
能实现芯片间的相互通信
It could help solve energy consumption issues in data centers
这有助于解决数据中心的能耗问题
and potentially bring down costs.
有可能降低成本
In these huge factories,
这些大型工厂中
processors are closely arranged together
处理器密密麻麻地排列在一起
so they can communicate at high speeds.
因此它们能高速交换信息
But this also generates a massive amount of heat.
但这也会产生许多热量
So not only do you have to put the power in
所以 你不仅要给
to power all of these components,
所有设备部件供电
but then you have to use an enormous amount of power
还要用大量电力
to actually cool them down
来冷却设备
to stop them literally melting.
以防熔化
What this company is trying to do
这家公司试图实现设备互连
is to come up with interconnects and ways of connecting components
用传输光而非电子的导线
that use light rather than electrons over wires.
实现设备组件之间的互连
So that means, in theory,
理论上意味着
that you could have a processor in one room,
可以在某个房间有一台处理器
memory in another,
另一个房间有内存
storage in another,
另一个房间有储存器
and that would help with the density
这有助于降低设备密度
and the power and the heat.
减少供电和热量
And that’s definitely a solution
这的确是许多公司
that’s being tried by numerous companies,
正在试图实现的解决方案
but we’re still waiting for that to show,
但我们仍在等待
you know, practical everyday results
能够日常使用的实践成果
and be widely adopted.
得以全面推广
三、神经形态计算 印度 孟买
Human brain is very efficient.
人脑很有效率
Our brain operates all the time.
一直在不停运转
It doesn’t dissipates more than 10 to 20 watts,
而其功耗不足10到20瓦
which is very small value.
这是一个很小的数值
To solve the problem of Rubik’s Cube,
为了解决鲁比克方块问题
so the amount of energy
与实际运行的系统相比
which is required to learn and then operate is very low
学习和转动魔方
as compared to the actual systems.
所需的能量值很低
Just solving the problem,
抛开学习不谈
put learning aside,
仅仅是解决这个问题
that itself will need thousands of processors
它本身就需要数千个处理器
to become parallel
并行处理
and then dissipate a lot of power,
消耗大量电力
which can go beyond megawatt.
能耗可能会超过兆瓦级
These scientists in India
多年来
have been trying to build low-energy chips
这些印度科学家一直试图制造出
mimicking the way the human brain works for years.
模仿人脑运作方式的低能耗芯片
The human brain is packed with neuron cells
人脑中充满了神经元细胞
that communicate with each other through electrical pulses
神经元通过名为“尖峰”的电脉冲
known as spikes.
相互交流
Each neuron releases molecules that act as messengers
每个神经元都会释放分子充当信使
and control if the electrical pulse
控制电脉冲
is passed along the chain.
是否沿着神经链传递
This transmission happens in the gaps between neurons,
这种信息传递发生在神经元的间隙中
which are called synapses.
这种间隙名为“突触”
Neurons and synapses,
神经元和突触
they are fundamentally different
它们与电脑的计算方式
from the way computers compute.
有着本质上的不同
Computers use transistors, which are binary.
电脑使用晶体管 是二进制的
They do zero, one, on, off.
编码为0和1 开和关
The brain sort of does a similar thing
人脑和电脑做的事有些相似
except that it works in analog.
除了人脑以模拟方式运作
It’s also partly stochastic.
人脑同样存在一部分随机性
A neuron may spike or may not spike
神经元可能产生 也可能不产生尖峰信号
depending upon inherent internal randomness.
这取决于神经元内在的随机性
The brain itself doesn’t like to do repetitive things.
人脑本身不喜欢做重复的事
It likes to do everything with slight variance.
而喜欢做存在细微差异的事情
This sort of enables creativity.
这能激发创造力
It sort of enables problem solving.
有助于解决问题
And the youngsters’ creative imagination
年轻人的创造性想象力
is not neglected.
不容忽视
Once there was a bear that was twice the size of the earth
从前有一只熊 它有地球的两倍大
and lived on the sun.
住在太阳上
The connections are different.
两者的连接方式不同
So if you look at a computer chip,
如果你观察一个电脑芯片
every transistor roughly talks to about 10 transistors.
每个晶体管大约和十个晶体管交换信息
Every neuron, on the other hand,
另一方面 每个神经元
talks to about 10,000 other neurons,
与大约一万个神经元交换信息
and that’s literally what makes it super parallel.
这正是人脑拥有超级并行处理能力的原因
Neurons use time as a token of information.
神经元用时间代表信息
Computers don’t understand time per se,
而电脑本身并不了解时间
they just sort of, you know, work with digital numbers
电脑只是在处理数字
and do mathematical operations.
进行数学运算
To mimic the architecture of our brain more closely,
为了更接近地模拟人脑构造
the team designed an artificial neural net
该团队根据人脑生物神经网络
based on the biological neural networks in our brain.
设计了一个人工神经网络
Each artificial neuron
每个人工神经元
is built with a few silicon transistors
均由几个硅晶体管构成
and connected to other neurons.
并与其他神经元相连
We try to mimic the auditory cortex.
我们试着模仿听觉皮层
So the auditory cortex has this neurons
听觉皮层的这个神经元
which are connected in random.
是随机连接的
And that sounds weird,
这听起来很奇怪
but it turns out it serves a very important function,
但事实证明在语音处理中
which is in speech processing.
它有一个很重要的功能
What we are doing is that there are recurrent connections,
我们所做的人工网络存在递归连接
which means that the neuron could go to any neuron
这意味着神经元可以连接包括自身在内的
including itself,
任何神经元
creating some sort of loopy pathways which are fairly random.
形成某种高度随机的回路
And these loops are part of the architecture,
这些回路是网络构造的一部分
they are called liquid state machine.
它们被称为液态机
They naturally are able to process time series,
它们天生就能处理时间序列
like when you hear speech,
就像你听到的语音
which is sort of occurs in time.
是在一定时间内发生的
We are doing some analog compute.
我们正在做某些模拟计算
There is some noise also in the neurons,
神经元中也存在某些干扰信息
so there could be some stochasticity.
因此可能存在一定的随机性
In the traditional or conventional computation platform,
在传统或常规计算平台中
everything is digital.
一切都是数字化的
Everything is defined based on logic one and zero.
一切都根据0和1的逻辑来定义
Analog computers,
模拟电脑
which are eventually signals which can have any value,
最终能摸拟任何有价值的信号
would be able to get more complexity
在操作中获得更多功能的同时
at the same time more functionality in the operation.
也会带来更多的复杂性
Here we have our neural network chip
这是我们测试板上的
which is on the test board.
神经网络芯片
So what we’ll do is we’ll speak to the computer,
我们要做的是与电脑对话
and that will sort of convert the speech into spikes.
它会把语音转换成尖峰信号
It’s gonna get processed in the chip.
芯片会对其进行处理
We run the code.
我们运行代码
One.

So it detects a one.
于是它检测到1个一
This neuromorphic chip
这种神经形态芯片
has been able to recognize speech
已经能够识别语音
using a fraction of the energy used by a traditional chip.
耗能仅为传统芯片的一小部分
Just like biological neurons,
就像生物神经元一样
the artificial neurons wait for inputs
人工神经元等待着信息输入
and only spike when data comes in,
只有数据输入时才会出现尖峰
and that means lower energy consumption.
这意味着能耗更低
Each neuron is also highly efficient
通过应用一种名为“带间隧穿”的
by using a quantum mechanics phenomenon
量子力学现象
called band-to-band tunneling.
每个神经元同样都是高效的
So electrons are still able to pass through transistors
因此 电子仍然能以极低的电流
with very low electric current.
通过晶体管
Quantum tunneling is nothing
量子隧穿对人类而言
that humans essentially have experience for.
根本没有经验可以借鉴
If you have a human being that has to cross over a hill,
如果一个人必须翻过一座山
the human being has to walk up the hill,
那么他必须消耗能量
burn energy,
爬上山
and then walk down, which again burns energy.
然后再次消耗能量走下山
And so there is no way for you to cross this barrier
所以你没办法在不消耗能量的情况下
without using up energy.
越过这个障碍
If you make this hill thinner and thinner,
如果你把这座山变得越来越薄
narrower and narrower
越来越窄
until it goes to the size scale
直到它的尺寸达到
on the order of, say, a few electron wavelengths,
比如说 大约几个电子波长的量级
at which point,
这时
an electron doesn’t need to go above the barrier
电子无需翻越势垒
to surmount it,
就能穿过它
it can, in principle, tunnel through it,
原则上 电子能穿过势垒
which means it will sort of disappear on one side
这意味着电子会在一侧消失
and appear on the other side magically.
然后神奇地出现在另一侧
If you take a transistor,
拿一个晶体管来说
there’s this same idea that when it turns on,
原理是相同的 打开晶体管时
we have a barrier that we sort of reduce,
我们相应减少势垒
and when we want to turn it off,
而当我们想关闭它时
we increase the barrier.
就增加势垒
That’s when you expect no current to go through,
这就是你预想的没有电流通过的理想状态
but that is not true by quantum mechanics.
但就量子力学而言 这不是事实
So even through a barrier, when it’s narrow enough,
只有垫垒足够窄时
electrons can sort of, quote and quote, tunnel through it.
电子才能“隧穿”通过势垒
So far, there are 36 artificial neurons on the chip,
截至目前 芯片上有36个人工神经元
and their energy consumption is remarkably close
其能耗与生物神经元
to biological neurons.
非常接近
If we want to approach the number of neurons
如果我们希望接近人脑中
which are already in human brain,
存在的神经元数量
so we need billions of them,
那么我们需要数十亿个神经元
and therefore we need to reduce the energy per spike
因此我们要尽可能减少单个神经元
generated by a single neuron as much as possible.
产生的单个尖峰能量值
加利福尼亚州 圣克拉拉
Intel, one of the world’s biggest chipmakers,
英特尔 世界上最大的芯片制造商之一
has been working on brain-inspired chips for seven years.
七年来一直致力于开发以人脑为灵感的芯片
They’ve released two versions of
英特尔发布了两个版本的
their neuromorphic research chip Loihi.
神经形态芯片Loihi
The latest one is only 32 square millimeters in size
最新版大小仅为32平方毫米
but contains up to a million neurons.
却包含多达100万个神经元
The company said Loihi has demonstrated
英特尔公司表示 Loihi实现了
multiple applications,
多种应用功能
including recognizing hazardous odors
包括识别有害气味
and enabling robotic arms to feel.
让机器臂有感知力
In the near term,
从近期来看
based on the most recent results we’ve had with Loihi,
基于Loihi的最新成果
I would say that solving optimization problems
我认为解决优化问题
are the most exciting application.
是最令人兴奋的应用研究方向
Things like planning, navigating,
比如规划 导航
you can think of it of solving the shortest path in a map,
你可以把它想象成解决地图上的最短路径
for example, if you want to get from point A to point B.
例如 你想从A点前往B点
These kind of optimization problems
这类优化问题
are conventionally really compute heavy,
通常计算量很大
but we found that on a neuromorphic architecture,
但我们发现 在神经形态构架上
we can reduce the computational demands
我们可以将计算需求
by orders of magnitude.
减少几个数量级
So over a 10 times speedup
速度提高了10倍的同时
as well as over a thousand times energy reduction.
能耗减少了1000倍
Probably the majority of the groups
或许大多数
that are working with Loihi
与Loihi合作的团队
sees robotics as kind of the long-term,
都把机器人科学当作神经形态芯片
best application domain for neuromorphic chips.
实现长期应用的最佳领域
After all, brains evolved in nature to control bodies
毕竟 人脑的自然进化是为了控制身体
and to deal with real-world stimulus,
应对现实世界的刺激
respond to it in real time,
实时做出反应
and adapt to the, you know, unpredictable aspect of the real world.
适应现实世界中不可预测的因素
And if you think about robotics today,
想一想今天的机器人科学
they perform very rigidly prescribed tasks.
它们执行着十分严格的规定指令
But if you want to deploy a robot into the real world,
但如果你想把机器人应用于现实世界
we really don’t have the technology yet
那么我们还未掌握足够的技术
to make them as agile,
让机器人能像我们希望的那样
as adaptive as we really would like.
灵活且有适应力
And so that’s where we see neuromorphic chips
这就是我们看到的神经形态芯片
really thriving.
得以蓬勃发展的地方
Intel and its collaborators
英特尔及其合作者
have used the Loihi chip to power iCub,
用Loihi芯片为iCub供能
a humanoid robot designed to learn
这是一款以儿童可能用到的方式
in a way that children might.
进行学习的人形机器人
This iCub robot is intended to interact with the environment
这款iCub机器人旨在与外界交互
as opposed to being pre-trained
而不是提前使用静态数据集
with static data sets beforehand.
进行预训练
So using kind of saccadic motions,
通过某种扫视动作
moving the eyeballs,
移动眼球
it can trace through and detect change
就能追踪并检测
or edges in the objects in front of this robot.
机器人前方物体的变化与边界
And using a fully neuromorphic paradigm,
通过一个完整的神经形态范式
we’re able to design a network
我们能设计一个网络
that understands and learns new object classes,
这个网络 即使在不同的角度和姿势下
even at different angles and different poses.
也能理解和学习新的对象类
It adds to its understanding,
这增加了机器人的理解力
it’s kind of dictionary of understand objects over time.
久而久之会变成一部类似理解物体的辞典
What is this called?
这叫什么?
That’s a mouse.
鼠标
Oh, I see.
哦 我懂了
This is a mouse.
这是鼠标
Let me have a look at that.
让我看看鼠标
Sure.
当然可以
Can you show me the mouse?
你能让我看看鼠标吗?
Here is the mouse.
这是鼠标
So far
截至目前
all of this technology is at an early stage.
所有这些技术都处于早期阶段
But to fight climate change
但是为了应对气候变化
and power the progress of civilization,
推动文明进步
we need at least one of them to be a breakthrough.
我们至少要实现其中一项突破
Greener chips absolutely have to happen.
更环保的芯片一定会出现
People are concerned about the environmental impact,
人们关注环境影响
people are concerned about the efficiency,
关注效率
people are concerned about the results
关注结果
and the way that chips can impact modern life.
以及芯片影响现代生活的方式
So there’s been,
所以一直存在着
you know, an absolute blossoming of investment
对芯片创业公司的投资热
in chip startups and trying these things.
以及与芯片相关的研究热
This is definitely the time of great ideas.
这一定是个百家争鸣的时代
Will they all be successful?
所有想法都会成功吗?
Absolutely not.
肯定不会
But now, I would say more than ever,
但现在 与以往相比 我更想说
there is a chance that out there
在世界的某个角落
in some corner of the world,
某家芯片初创公司
there is a chip startup that will be everywhere
可能会在接下来的10年内
within maybe the next 10 years.
占领全球市场
You can’t get your enjoyments
你不能只从科学的成就中
from only the successes in science.
汲取快乐
You need to get your enjoyment
而要享受
from just enjoying solving hard problems.
解决难题的乐趣
And so if you fail,
如果你失败了
that is just part of the process
这只是最终找到
of eventually finding a solution.
解决方案过程中的一部分罢了

发表评论

译制信息
视频概述

芯片安全已上升到国家安全战略层面,谁掌握了芯片技术,谁就掌握了科技发展的主动权。如何突破技术瓶颈?如何大幅提升芯片处理速度,同时降低能耗?如何制造出完美的电脑芯片?碳纳米管以及神经形态计算方式或许是其中的关键。

听录译者

收集自网络

翻译译者

Kris

审核员

审核员SRY

视频来源

https://www.youtube.com/watch?v=VsUF_CBJq50

相关推荐