From a hardware point of view, the biggest difference between the three lies in the way AI operations are implemented. We know that the difference between AI calculations and common computational tasks on mobile phones lies in achieving machine learning. Reasoning for deep learning takes a long time. Keeping the high speed, a large amount of calculation processing. The increase of the peak data throughput, the increase of the calculation time length, including the increase of the calculation speed requirement, have all made new demands on the computational power of the AI chip.
The difference between these three chips lies in how to achieve stronger computing power.
Among them, Unicorn 970 uses a 'hard implementation' approach, adding a new computing unit outside the traditional computing unit such as the CPU, GPU, etc. - embedded neural network processor NPU. Just as the GPU is dedicated to image operations, The NPU is also dedicated to convolutional neural network calculations to achieve machine learning related functions. The same AI calculation task will have about 50 times more energy efficiency and 25 times more performance than a simple CPU operation without the NPU. Faster, and lower energy consumption, so as not to affect the normal use of mobile phones.
Different from the hard-implementation mode of the Unicorn 970, the Apple A11 adopts the 'soft implementation' mode, and does not have an independent neural network processing unit. However, the AI is adapted to self-developed GPUs, self-developed ISPs and DSPs and other conventional hardware modules. Adjustment, integration of the 'neural network engine' greatly enhances the processing speed of picture calculations. In AI tasks, many involve the learning and inference of picture content, such as AI beauty, etc., which enhances the ability of image computing. Naturally, Can improve the efficiency of AI calculations.
The concept of the AI Engine (Artificial Intelligence Engine) introduced by the Xiaolong 845 is similar to that of the Apple A11. It is also the adjustment of CPU, GPU, DSP and other hardware modules to replace the NPU's capabilities.
So, can the NPU unit in the chip really be replaced by other solutions?
To answer this question, we must first know what the AI chip can do.
Before having an AI chip, the opportunity to take advantage of AI tasks, most of the cloud computing methods adopted - upload data to the cloud, reasoning and then download the results to the local. This way of computing once is not only inefficient, but also It is easy to cause the application to be stuck.
But with the ultra-high-powered AI chip, the mobile phone can finally realize the end-side AI. Regardless of whether there is a network connection to the cloud, it can perform real-time, unconditional high-speed operations locally on the phone.
This kind of end-to-end AI brings four capabilities to the mobile phone, namely perception, awareness, motivation and security. Next we can discuss whether or not the NPU has produced any AI for the mobile phone AI from the implementation of these four capabilities. Impact.
How efficient is IntelliSense? Instantly smiles
How does a mobile phone perceive everything?
In the past, the user's information was decoded by means of input methods such as touch screens. At the same time, the GPS, Gyro and other two layers of hardware and software sensors were used to understand 'I'm where I'm from the ground' and so on.
However, with the increase of computing power, some algorithms that are originally in the cloud are entering the terminal, allowing the mobile phone to perceive more. For example, in image processing, in the past, mobile phones simply took photographs. Now, with increasing computing power, the mobile phone is also locally The image recognition model can be deployed. In the viewing stage, the mobile phone can know whether the pet or human is in front of the camera.
However, with the existence of the NPU, the overall improvement in mobile computing power is no longer limited to image processing capabilities, and the perceived information is more comprehensive.
For example, the Glory V10 has a smart music recommendation function. With the support of AI computing power, the user's motion frequency can be monitored in real time through the somatosensory capabilities of the mobile phone, thereby recommending music with similar melody and sports frequency.
Even the most widely used AI camera, there are many differences. Like millet Mix 2S and OPPO R15 have proposed scene recognition to distinguish the subject of shooting and the corresponding post-processing. But Huawei's P20 series can be in the end Side to achieve expression recognition, automatic capture to the subject's smile.
The reason is that smiles and fluctuations in the frequency of movement are often fleeting. As an independent computing unit, the NPU is solely responsible for AI calculations. It can call up resources as quickly as possible to complete processing. The CPU+GPU solution mode can also take care of mobile phones. With all of the computing tasks, it is possible to queue up. The people in front of the camera are laughing and the computing resources have not been scheduled.
With accurate knowledge, Wisdom recognizes screens and allows services to reach users
The so-called cognition is the process of processing information on mobile phones.
With the strong power of end-to-end AI, mobile phones are gradually moving from processing data to processing knowledge. For example, in the past, the iPhone's album division was divided according to the shooting location and shooting time, but now it has been able to achieve the cat, sea water The way to distinguish the content of photos. The phone can only understand the time from the past, GPS positioning these structured data to understand the unstructured 'knowledge' of cats and dogs and flowers.
In the EMUI, we can also see a function called Smart Screen Recognition. As long as the two fingers long press the screen to achieve a common reading of pictures and text, and through the AI to understand the semantics of the picture and text, showing the address , Movie Ticketing, Hotel Public Reviews, etc. Service Cards. Clicking on these cards can jump directly to third-party apps, greatly reducing the manual scheduling between App.
The emergence of this function is due to the fact that NPU's more powerful computing power can realize multi-dimensional processing of information: It can not only deal with natural speech and pictures at the same time, but also can be applied to jump through such active services by understanding the user's behavior habits.
Power Commander, knowing that users have the wisdom to save energy
A common understanding is that a lot of computing is very power-consuming. Otherwise, Bitcoin miners will not be looking for cities with low electricity prices. But on the power system of mobile phones, AI calculations can instead play a positive role.
As mentioned above, AI allows mobile phones to have more powerful cognitive abilities. In the mobile phones that have NPUs, this recognition is happening at every moment, and finally realizes a kind of personal habits of mobile phone users. In-depth understanding. Take the battery life, equipped with a unicorn 970, Huawei and Glory applied EMUI have joined the smart power-saving features. That is, through the individual user of the application of the habit of deep learning, remember which applications are often Used to predict which applications are to be opened in which scenarios. Under this system, resources are provisioned again. With independent computing units, mobile AI is learning and calculating all the time, 'cutting off' those enabled. Low-powered app's power supply, so that the phone can get better endurance.
Not only battery life, but also the wisdom of smart phone scheduling is also playing a role in the Android mobile phone system is not smooth. In ordinary mobile phones, each application is queuing to get their own operating resources, the long queues become longer and longer. It's also getting worse. The NPU is like a commander, judging how much resources they need in the day-to-day performance of the application, and making the queues become real-time on-demand assignments.
In this way, not only the utilization of memory and computing resources of the mobile phone application is improved, AI can also remind users regularly, which applications have traffic changes, and which applications have not been started for a long time. This helps the user to clearly understand the internal operation of the mobile phone. Condition. According to the data of Glory V10, in this mode, the system scheduling performance increased by 60%, and the operational fluency increased by 50%. From the power battery to the memory, the overall power system efficiency of the mobile phone has been improved.
Permanent barriers to security,
There is an end-side calculation that dares to hand over the data to the phone
In the era of cloud computing, user privacy issues are like a high wall. To obtain better services, data needs to be uploaded to the cloud for processing. However, once the data is uploaded to the cloud, it may be revealed. However, the existence of end-side AI perfectly avoids this problem. The calculations are performed locally. There is no personal privacy exposed in the common cloud space.
After joining the NPU, you can provide more services for the user on the premise of ensuring privacy. The understanding of the user's behavioral habits is completely carried out in a closed space without a net, and the mobile phone can be more 'understand'. Current HUAWEI P20, Glory V10 can already do intelligent adjustment of screen brightness, background resources, debris files, etc. In fact, the mobile phone also knows the driving route you use most often, the itinerary for the next few days, where you like to go, etc. More information, this information on the cloud will certainly feel offensive, but in the face of AI's blessing of information security, the phone will become a qualified personal assistant.
For example, when a user subscribes to a restaurant through voice assistance, the mobile phone can directly recommend a suitable route through the user's travel habits, and even can recommend the food to the user. Such an intelligent experience based on a sense of security is likely to make the user unknowingly. Reliance on AI phones.
Seeing here, we will find that NPU is not just a simple technology addition to mobile phones, but enhances the user's sense of use of mobile phones through the subtleties of perception, awareness, motivation, security, etc. Users may not Understand the so-called deep learning, but you will surely feel that the mobile's endurance ability is getting stronger and stronger, and it is more and more in line with your own habits.
Now that mobile AI is still in its infancy, it is very difficult for users to form a precise understanding of AI mobile phones. At this moment, there is a window period, from the red rice's thousands of AI series to the intelligent photography of OPPO. AI-capable mobile phones can be called the AI's banner 'hot'.
When users actually use these mobile phones, they can't realize the mobile phone's learning of their own behaviors, but they can't realize the 'AI filter' that is the same as Mito Xiuxiu. For the operating system speed, the battery life is improved. AI is not AI, it has no change in the sense of use. For the mobile phone product, the photographic effect, endurance, and ease of use are all eternal monetary systems. It is consumers who affirm their value. The only reference.
The window period is always short-lived. It is better to grasp a technology that can truly change the product experience.