A key with three doors: AI chess game behind the unicorn 810

Until now, mobile phones or need AI, is no longer a problem.

After Since 2017, Huawei kicked off the side of the AI ​​with Kirin 970 chip, we have witnessed the AI ​​to bring the phone to a lot of imagination. AI mobile phone photography, AI camera acclaimed; intelligent voice assistant to the scene pulled rose to a new height; a lot of imagination to open up a new mobile ecosystem based on the ability of third-party development and application of AI. It can be said that Huawei mobile execution AI AI accelerate the road side has made it clear in the end. But then how to walk the road, it has been widely outside attention. We are cautious at every step, or march Warp?

June 21, in Wuhan nova 5 series new conference, Huawei released a new mobile phone chip AI - Kirin 810, which is Huawei's first self-developed mobile phone NPU architecture Leonardo da Vinci AI chip.

A key with three doors: AI chess game behind the unicorn 810

While this chip from the information contained in, Huawei is almost certain policies on the movement of AI, is still maintained a high-speed running, many-fold for the overall development and rapid integration of AI.

I say this is because not only are they a unicorn 810 equipped NPU chip, or a key to open multiple industry boundaries. DaVinci architecture and layout of the formal merging AI movement, not just change a chip and a cell phone products, but the overall Huawei AI chess Road to accelerate, the Pratt & Whitney AI's strategic goal was another step upgrade.

AI strategy behind Kirin 810, the keyword is "pass."

When in the sky: Huawei AI chess game finally merge

Last year, HC conference, Huawei has released one of the most important basic technology for the AI ​​era - Leonardo da Vinci architecture and using this chip architecture of AI 310/910 rising. This time we know that the use of the Leonardo da Vinci Kirin 810 in the NPU architecture, but the average reader and may not clear the relationship between the two.

In simple terms, this means that Huawei layout respectively in terms of accelerating the underlying technology and AI AI side, completing a victory rendezvous.

In October last year, Huawei was once very famous Leonardo da Vinci architecture.

A key with three doors: AI chess game behind the unicorn 810

这种华为自研AI芯片的底层架构,之所以名为达芬奇,是因为达芬奇是人类历史上在“全才”领域的代表人物。“全才”,正意味着AI所需的计算特性。AI任务运算的难点在于必须对多种计算技术和AI算法进行统筹把握。不同于经典计算,神经网络任务普遍需要张量计算模型来保证处理效率,而从本质上来说,张量矩阵运算是一种要求大吞吐量、低延迟的计算模式。而这其实契合了很多网络运算的需求。因此AI到来之际,全面布局多个领域的“全才”华为,恰好具有独特的优势。

以高性能3D Cube计算引擎为基础,达芬奇架构可以针对张量矩阵进行加速。其效果已经在昇腾系列芯片的应用中得到了体现,并布局到了云端与边端侧AI加速等计算场景。从达芬奇架构衍生出的AI底层技术体系,可以看作华为在AI领域布局的基础设施。

这种3D Cube计算引擎已成为整个端侧AI行业的发展趋势。回顾行业内前几代AI手机芯片,最初大多采用DSP来完成基础的scalar (标量)运算,后来逐渐出现加入能够处理vector(向量)运算的部件,一路发展至今,NPU已成为AI运算的标配单元。随着AI运算的数量和复杂程度激增,下一阶段,更具有挑战的Cube运算将成为端侧AI的全新高地。

在这方面,麒麟芯片始终引领着行业趋势,无论是最早在业内集成独立NPU,还是此次率先采用自研达芬奇架构,都是在端侧AI领域的一次全新突破。这种高性能的3D Cube计算引擎,已经初显其AI实力,根据发布会介绍,达芬奇架构NPU拥有业界最优算力和最高能效,实现单位面积下的最高算力,FP16精度和INT8量化精度业界领先。

而另一方面,早于达芬奇架构的提出,华为就已经在麒麟芯片上探索基于端侧硬件进行AI加速的形式。基于产业合作的模式,将第三方IP的NPU引入麒麟970、980系列芯片的SoC,经历了对工程化技术的验证、产品形态的发展、软硬件开发生态的建立,移动AI这条路已经被证实了可行性。

或许可以这样理解,麒麟芯片、HiAI平台是一支轻骑兵部队,需要去抢占移动AI的先头高地。而达芬奇架构、昇腾芯片等基础能力是华为的AI大本营。在一场行动开展初期,大本营和先锋营必须分头行动,各战所长。而在轻骑突进阶段结束后,两军会师的“军团战”才刚刚开始。

A key with three doors: AI chess game behind the unicorn 810

会师之日,就在麒麟810发布的这一天。目前来看,达芬奇架构集成在NPU之后,初试身手效果不俗。在AI-Benchmark跑分中,麒麟810以超过3万分的成绩力压于骁龙855的25428分。

当然这里还要解释另一个问题,面向未来的AI产业发展,华为这“两军打通”,进行统一架构到底目的何在?总体而言,有三个能力是华为AI会师后一定会出现的:

1、全栈打通,在终端产品上更好执行深度AI开发,让底层AI技术具备定制化和与深度可控性。使用自研AI架构之后,华为手机将可能探索更多深度AI应用,并且降低开发成本,缩短开发周期。

2、执行云、端一体化与高效配合,产生更流畅的AI产品布局。我们知道,移动AI更多时候应该是云端合作的。云侧处理大型AI任务,端侧执行实时化与涉及隐私数据的本地计算。而更好的用户体验,显然是云与端无缝合作的AI加速模式。当华为的云端AI与端侧AI计算架构打通,可以提高华为手机背后云端一体化AI任务的兼容性,让这类任务更加快速流畅地落地。

3、通过AI,激活华为手机与全场景设备的连通性。需要注意的是,基于达芬奇架构的昇腾系列芯片,还有昇腾lite、昇腾nano等型号。这些芯片的特点是功耗更低,显然目的是进入AI+IoT这个新领域。换句话说,未来IoT设备也会纳入达芬奇架构家族。这也是华为AI全栈全场景战略的一部分。

万物AI的开始,手机是天然枢纽,而统一架构可以更好处理任务。华为电脑与手机之间的“一碰传”功能就要基于大量的AI算法来实现。当达芬奇架构在各品类华为终端设备,甚至在华为全场景生态设备间打通,那么其中可以迸发的技术想象力或许将是难以预计的。

麒麟芯片与达芬奇架构的融合,是注定会发生的事,而其发生在今天,还有一个价值在于天时。目前,华为在两大AI方向上都保持着领先性,合流可以形成1+1大于2的领先优势。目前能够具备从AI底层架构,到云端各领域布局,再到完整SoC芯片产业链,全球产业链中只有华为。于是天时在我,机法随心。

中通地利:nova 5踏入端侧AI序列后的产业轨道

麒麟810芯片另一个打通的,是AI能力在华为手机产品体系中的拓展。

产品与技术形态,就是手机的战地。近两年,旗舰技术的下放与普惠化,是华为手机地利优势的核心保障之一。而麒麟810展现出强大AI能力,意味着在nova 5的核心竞争力极大幅度增强。Nova 5发布后,很多媒体评价说它是旗舰的配置和技术,中端机型的架构。这个技术差异的思路应该会继续在华为手机中发展,而芯片端的技术差异化是这一战略的基础。

所以麒麟810打通AI与nova系列的联系,证明了华为此前披露将让核心技术优势普及到旗舰以外系列的所言不虚。麒麟810不仅让7nm制程这个旗舰机芯片的标配推进到了nova系列,同时新的强大NPU,将保证nova 5具备旗舰级别的AI能力,以及享受未来多元化的AI开发生态。

A key with three doors: AI chess game behind the unicorn 810

AI走向nova,更重要的价值在于,侧面证明了在华为的理解中,AI不是旗舰机才有的特殊优势,而是未来每部手机中的基础能力。当所有机型,甚至整个产业都能应用端侧AI能力,移动AI才有可能引发像智能机革命一样的浪潮,移动AI将逐步走向真正意义上的普惠化。

打通AI与华为更多机型的界限,确定未来AI成为基础能力的产业轨道,这是麒麟810打通的第二道门。

外通人和:开发者的AI捭阖

打通达芬奇架构,联通新产品序列,这“两通”的直接作用,是麒麟810可以进一步打通AI与开发者的界限。

所谓人和万事兴,我们都知道千万开发者的智慧是AI时代的基石,那么麒麟810对AI开发者生态的影响是什么呢?我们可以从近到远,看到三个改变已经开始。

1、更低的AI开发技术门槛。

从技术基础上看,打通达芬奇架构和新的芯片设计,让麒麟810算子支持数量达到240个以上,让AI加速能力具备更好的通用性。帮助AI开发者获得更好的开发兼容性,降低开发门槛。

在开发工具上,打通达芬奇架构后,HiAI移动开放平台将获得华为自身开发工具链的更好支持。从框架、开发平台到工具组件,华为的AI开发技术将赋能移动AI开发者。搭配强大的AI算力,提高开发者的开发效率,降低开发成本。如今AI技术日调用量已经超过30亿次的HiAI,这次可谓是在基础层进行了革新和洗礼。

A key with three doors: AI chess game behind the unicorn 810

2、更大的目标受众。

麒麟810进入AI序列,另一个带给开发者的直接改变,是移动AI应用的目标用户群更大了。这对于开发者来说当然是好消息。而达芬奇架构与麒麟芯片的打通,同时也意味着开发者将可以更好地执行云端一体化开发,以及IoT与手机进行AI联动的尝试。这对于开发者来说都是宝贵的市场机遇。正所谓广泛市场,是最基本的行业普惠。

3、更明确的市场潜力和前行方向。郑州不孕不育医院:http://byby.zztjyy.com/yiyuanzaixian/zztjyy//

麒麟810对于开发者们来说,不仅意味着技术与市场上的迭代。更是一个信号传递器,它的出现,意味着华为全栈全场景AI战略正式与手机端接通,那么移动AI开发者的全栈全场景AI开发也可以顺势启动。开发者接下来将享受更低的AI开发难度,云端一体化以及AIoT带来的市场机遇,以及多种使能方式结合的华为AI开发生态。先有联接,再有普惠,在AI道路上同样如此。而AI能力在华为手机序列中的拓展,则意味着移动AI通用化趋势开始行动。AI开发者未来将是面向所有用户的开发者,这点对于很多开发者来说将是定位的改变。

从技术,到市场,再到战略判断,麒麟810打通了AI开发者一直关心的重要节点。开发者可以纵横捭阖尽情挥洒脑洞,是普惠AI唯一的未来。

结束语

Pratt & Whitney AI, AI democratization, or called people-oriented AI technology, the overall direction is one of the core concerns of the past two years AI industry, but also governments, academic and research institutions overall vision for the future technical development of AI.

The birth of a technique called Pratt & Whitney in the end how? This question has a strong practical orientation. Only the underlying technology has a universal value and low cost characteristics, a sufficient number of products and can avail themselves, a technique in order to be judged as Pratt & Whitney.

A key with three doors: AI chess game behind the unicorn 810

It may be noted that the combination of mobile phone and AI if expectations of Pratt & Whitney, the emergence of Kirin 810 and this should be just a few trends:

Leonardo da Vinci self study architecture in the mobile phone chip, Huawei can mean better technology to control costs, reduce the technical difficulty through continuous investment in research and development, improve the technological generalization. Pratt & Whitney grasp from the source to ensure the ability.

The AI ​​chip into the nova series, which means there will be more products which began accommodate AI; for blessing HiAI platform means more developers will be able to apply AI techniques.

From the technology, products, people in these three areas, Pratt & Whitney AI 810 are core objectives behind strategy unicorn. While this chip is like a key to the door to many AI before quietly locked open. Only more keys to open more of them, we will eventually usher in a barrier-free AI ecology.

Perhaps one day, everyone can become a Leonardo da Vinci.


Guess you like

Origin blog.51cto.com/14393777/2413282