Zhi DongXi (public number: zhidxcom)
Author | Yunpeng
Edit | Desert Shadow
Nowadays, everyone may be able to talk about the term "AI" twice. All kinds of intelligent hardware devices continue to be deeply involved in our lives, from smart cars to small mobile phones.
Behind the outbreak of intelligent hardware, AI algorithms are also developing rapidly, algorithm models are becoming more and more complex, and the requirements for hardware are higher, and the concept of "AI chips" has become one of the focuses of the AI industry in recent years.
In fact, modules that can be used for artificial intelligence applications and achieve AI acceleration can be called AI chips in a broad sense. For example, the NPU (neural network processor) module in most of our mobile phone processors is an AI chip.
Just two days ago, OPPO released its first self-developed chip, which is an NPU chip dedicated to image processing, called Mariana X, which has the same name as the world's deepest trench, and OPPO wants to go to the deepest water of the self-developed chip.
With the new DSA architecture for AI applications, TSMC's 6nm advanced technology and 40 times the ai energy efficiency ratio improvement, Mariana X has pushed mobile phone computing images to the era of "4K+20bit RAW+AI+Ultra HDR", and at the same time, for the first time, the AI imaging algorithm has been vertically integrated with sensors and SoCs, completing the closed loop of the image link.
In a word, Mariana X, as the standard chip for developing a dedicated AI algorithm and hardening the full-link vertical integration capability of the computing power, bridges the gap between THE AI algorithm and the hardware, and becomes a "bridge" to accelerate the landing of the AI algorithm.
As OPPO founder and CEO Chen Mingyong said, OPPO has entered the "deep water area" of technology research and development, for OPPO, the landing of Mariana X proves the technology and capabilities of OPPO's self-developed chips, and for the AI industry, Mariana X may provide new ideas for the accelerated landing of AI in the mobile field.
First, the AI energy efficiency ratio has soared by 40 times, the algorithm landing video has become possible, and the DSA architecture refers to the calculation of images
From the perspective of chip technology architecture, common AI chips mainly include general-purpose chips (CPUs, GPUs), semi-customized chips (FPGAs), fully customized chips (ASICs), brain-like chips and software-defined hardware chips.
Mariana X NPU is mainly used in the field of mobile phone image computing, relatively more specific applications, and the chip is specially designed to enhance the image experience of OPPO mobile phones and customized dedicated AI chips, which are close to the assic chip and bring super high efficiency, while having the flexibility that AISC does not have.
At present, the general mobile phone SoC contains ISP (Image Signal Processing) module, as the name suggests, the main role of the ISP is to process the information collected by the image sensor, it is also an ASIC chip, but most of them are integrated in the mobile phone SoC.
The core difference between is the ISP and the Mariana X is that the ISP does not have AI performance, and it must rely on another AI computing unit to cooperate in the SoC. For example, the ISP of Apple's A-series chips also needs to rely on the entire chip's Neural Engine to provide AI computing power support.
Mariana X's AI computing power has reached 18TOPS, which can already be compared to Apple A15, and while the peak computing power is strong, achieving better energy efficiency is more important to OPPO, which is particularly important on mobile devices such as mobile phones.
The Mariana X is capable of achieving an 11.6TOPS/W energy efficiency ratio, a rare two-digit pair of energy efficiency performance in mobile devices, meaning that the phone can handle larger amounts of data under the condition of consuming a given amount of power. It is worth mentioning that the results are based on the real commercial complex AI algorithm and randomly generated pictures, in fact, if based on real-world portraits, landscapes and other real pictures for processing, Mariana X's energy efficiency ratio will be higher.
Compared with the Find X3 Pro equipped with the Snapdragon 888, the performance of the Mariana X running OPPO's self-developed AI noise reduction model can reach 40FPS, which is about 20 times that of the OPPO Find X3. On this basis, the power consumption of Mariana X is only 800 mW, which is only half of the Snapdragon 888 on oppo Find X3, so the energy efficiency ratio has reached 40 times, which is undoubtedly a "qualitative variant".
In fact, although mobile phone computing photography has developed for many years, most AI algorithms can still only be processed for a single or multiple photos, such as AI beauty, AI night scene photography post-processing, essentially still time for calculation, if you want to calculate more, it means that consumers often need to wait longer after shooting processing.
However, such an algorithm is difficult to apply in the face of the need to process a large number of video shots in real time. Because in video shooting, the minimum performance threshold for AI algorithms to run must also have 30FPS, which means that the chip needs to provide more than 30 times the performance in order to allow the post-processing algorithm to run in real time.
It can be said that Mariana X makes the application of the algorithm in the video side from impossible to possible, which is "from quantitative change to qualitative change" for the improvement of the application effect of AI algorithms.
Based on this AI performance and high energy efficiency ratio, Mariana X has achieved many breakthroughs in the field of mobile phone imaging.
For example, in terms of the dynamic range of imaging, Mariana X can support up to 20-bit Ultra HDR imaging, covering a maximum brightness range of 1 million:1, which is comparable to the "human eye level" and is also 4 times the mainstream mobile SoC HDR capabilities such as Snapdragon 8 and Tianji 9000 in the industry.
According to the comparison of the shooting examples given by OPPO, it can be seen that in the two types of scenes of portrait shooting and still life shooting, the video taken by the model using the Mariana X scheme can significantly improve the details of the shadows, and there is no overexposure of the highlights.
Behind the strong AI performance and intuitive improvements to the image experience, the Mariana X chip has also made a bold breakthrough at the architectural level.
Mariana X adopts the DSA architecture concept of the AI era, in short, DSA is "application in a specific field of architecture", Mariana X's "specific field" is obviously a mobile phone computing image.
At present, there are also modules for improving AI computing power in the general mobile phone SoC, different manufacturers may be called differently, but it can be widely regarded as a general NPU, which takes into account the algorithm model in many different scenarios.
In contrast, Mariana X's DSA architecture is more specialized, hardening the corresponding operators for OPPO's self-developed AI model, which is also one of the core differences between OPPO's image-specific chip and traditional NPU.
It can be seen that OPPO is bridging the gap between traditional hardware and AI computing power through a "dedicated core" chip design idea. In the core field of AI performance, Mariana X has enough ability to break through the "ceiling" of the current mobile computing photography.
Second, under the ai empowerment of thousands of industries, OPPO has become a dark horse in the AI track
The power point of Mariana X is the computing image of mobile phones, and looking at the entire AI industry, the tide of AI landing has penetrated rapidly in thousands of industries.
The prediction that AI chips will be everywhere is becoming a reality. From smartphones to smart homes, from data centers to edge devices, our lives are being fully invaded by AI chips.
In the field of smartphones alone, AI technology has penetrated into mobile games, visual recognition, audio interaction, communication optimization, intelligent charging and other corners.
The industry usually divides the AI industry into application layer, technology layer and basic layer, in fact, OPPO is currently facing the "3+N+X" strategy of the era of mutual integration of all things, including software and hardware basic technology, AI technology and differentiated technology represented by charging and imaging, in the application layer and technology layer of the AI industry, OPPO has more reserves.
The Mariana X is obviously a bold exploration and forward-looking layout of OPPO for the field of AI chips, and AI chips are also an increasingly important part of the current AI industry foundation.
In the field of general-purpose CPUs and GPUs in AI chips, GPUs can be used as AI accelerators because they are good at highly repetitive massively parallel computing tasks.
In this field, NVIDIA is leading the AI training chip market because of its mature and efficient GPU solutions, and its various AI accelerator products are widely used in the cloud and terminal side in the industry.
▲NVIDIA A100 GPU
Head players such as Nvidia, AMD, and Intel have become the biggest beneficiaries of infrastructure construction in the global AI field by providing AI accelerator solutions through customized CPUs and GPUs, but their products rarely involve mobile terminals, mainly for desktop platforms.
On the mobile side, the mobile phone AI chip is still mainly present in the form of NPU modules in the general mobile phone SoC, which is also closely related to the computing power requirements and power consumption requirements of mobile phones.
Due to the higher energy efficiency ratio requirements and the more "thousand faces" of customization requirements, mobile terminals also need targeted AI chip products to appear.
From the perspective of the past two years, the mobile phone AI chip market pattern is basically stable, with apple, Huawei, and Qualcomm firmly controlling the high-end market, and mediatek, ZhanRui and other manufacturers have performed well in the mid-end and entry-level markets.
But in fact, as mobile phones become more and more functional, the application scenarios represented by computing images are becoming more and more demanding for computing power and power consumption, and these general mobile phone AI chips have been somewhat stretched to cope with the AI needs of a specific field.
At this time, OPPO has undoubtedly become a new player in the field of mobile phone AI chips, with three years, after the efforts of thousands of R & D teams, finally brought the Mariana X image-specific NPU to our vision, and will soon land in the flagship Find X series next year.
In addition to the breakthrough of AI chips at the basic layer of the AI industry, OPPO has also grasped an important trend change in the AI industry, that is, the deep combination of software and hardware, writing algorithms into hardware, and doing "hardware-level algorithms", which has always been the goal that OPPO is committed to achieving.
On the application side of terminal AI chips in the field of mobile, OPPO has taken the lead in achieving breakthroughs in specific fields.
Rather than saying that OPPO is making a dedicated NPU chip for imaging, OPPO is using a new idea to achieve the accelerated implementation of AI algorithms on the terminal side.
Conclusion: OPPO Bright Sword AI chip industry
With its strong AI computing power and AI energy efficiency ratio, Mariana X has significantly improved the experience of mobile phone computing images, and also accurately cut into the AI chip track with the idea and method of deep combination of software and hardware and dedicated cores, becoming a new player in the terminal AI chip industry.
While AI is invading and subverting various traditional fields, AI chips are staging a new competition for dominance in cloud computing, mobile phones, security surveillance, smart homes, automatic driving and other scenarios. The battlefield pattern of terminal AI chips is complex, with new and old players and cross-border players blending and showing their own glory, and OPPO has also embarked on a path of its own chips.