Ten years ago, the British “Economist” used industrial electricity consumption as the main indicator to evaluate China’s GDP, but now computing power has become a new indicator. It seems that every business, every city is trying to increase computing power.
A reader left a message in confusion. Everyone said how many FLOPS their computing power can support this big model and that big data. Does it mean that the larger the number, the stronger the power?
Really not.
To measure the level of computing power, in addition to the number of calculations, it also depends on the accuracy of computing power, that is, how high the data accuracy and computing complexity can be supported.
Dismantling the problem of “computing partial science”, where will the intelligent computing center go next?
If the number of operations (FLPOS) represents the value of internal force, then the accuracy of calculation power is like “skills”, which determines whether the internal force can be used correctly and well. In a martial arts duel, someone slashes wildly with a dragon-slaying knife, and Linghu Chong relies on Dugu Nine Swordsmanship to hit every hit. Who do you think is more skilled?
Different computing tasks have different requirements for computing power, and appropriate “skills” are needed to maximize the value of computing power.
Specifically, according to different data types, the accuracy of the adapted computing power is also different:
For example, scientific calculations, weather forecasts, computational chemistry, molecular models, and astrophysical simulations require high data accuracy and require double-precision computing power (64-bit, FP64), which is a general-purpose computing power provided by supercomputers.
Dismantling the problem of “computing partial science”, where will the intelligent computing center go next?
AI model training, autonomous driving, smart city, AIGC and other businesses need to learn a large amount of data to train a complex deep learning model, and processing large-scale floating-point data is more suitable for single-precision computing (32-bit, FP32) , Half-precision computing power (16 bits, FP16). In recent years, large pre-training models have exploded, and waves of large models such as DALLE, ChatGPT, and Zidong Taichu have emerged, with parameters reaching hundreds of billions of trillions. Efficient training of large models requires a large amount of single-precision computing power.
The trained and deployed model only needs to infer various conclusions based on the input data in actual application, such as face recognition, license plate recognition, voice recognition, etc. This AI reasoning process deals with integer data, which is more applicable For integer computing power (INT8).
Dismantling the problem of “computing partial science”, where will the intelligent computing center go next?
So here comes the problem. Generally speaking, in a certain area, there are not only high-performance computing such as high-tech scientific research, smart cities, and autonomous driving, which bring the demand for general computing power; but also AI model training and AI applications. Reasoning, etc. to support industrial data-to-smart transformation requires a high demand for dedicated computing power.
If the computing power configuration in the region is not diversified enough, and a certain computing power is missing, it is equivalent to waiting for users to go to the battlefield to fight against the enemy, only to find that there is no way to use it or half the effort, which will inevitably limit the development of local digitalization.
Therefore, when constructing intelligent computing centers in various places, in the long run, it is necessary to consider the diversification and universality of computing power.
However, the reality is that the intelligent computing centers in many places are in the stage of “first approval and quick construction” and fighting on their own. Due to the rapid development of the AI industry/scientific computing/industrial digitization, there is an urgent need to supplement some special computing power to deal with computing power. The power is thirsty, and there is a lack of overall planning for general computing power.
With the advancement of a series of measures such as the East-to-West Computation Project and Digital China, entering a new period of development during the “14th Five-Year Plan”, the generalization of dedicated computing power has become a top priority that hinders the development of the local digital economy.
Dismantling the problem of “computing partial science”, where will the intelligent computing center go next?
The “Intelligent Computing Center 2.0 Era Prospect Report” released not long ago also emphasized that the current computing power configuration selected by individual regions has a certain “partial” phenomenon, which can only meet the needs of some subdivided scenarios, and cannot take into account multiple industries, Multi-field demand for integrated computing power. It is necessary to promote the integration of general-purpose computing power and special-purpose computing power, and drive applications to go deeper.
For various computing power users such as universities, scientific research institutions, enterprises, and governments, what benefits can computing power integration bring?
How much is the value of computing power fusion?
In a word, fusion of computing power means that users can take and use different computing power resources at will, whether it is the general computing power required by office digitization, or the dedicated computing power required by AI applications, or weather forecasts and biological predictions. The required high-performance computing power can be integrated, learn from others’ strengths, support the development of one’s own business, and become a “generalist” of computing power.
From this perspective, the integration of general computing power and special computing power will bring three obvious values:
One is the sustainability of the digital economy. The digital economy has become the main theme of the development of various regions, and the computing power application scenarios included in it are very extensive. The computing power is deployed according to the application scenarios to establish advantages faster. For example, a provincial capital city hopes to create an artificial intelligence highland. At the same time, universities take the lead in establishing remote sensing industry clusters. The integration of dedicated computing power and general computing power can support more abundant application scenarios and provide computing power for the long-term sustainable development of the local digital economy.
Dismantling the problem of “computing partial science”, where will the intelligent computing center go next?
The second is the reduction of comprehensive cost. The construction cost of computing power infrastructure is extremely high. Doing a good job in the configuration of various computing power in the early planning can improve the utilization rate of the infrastructure, ensure sufficient computing power, and at the same time leanly meet the needs of various tasks, thereby improving computing power. The comprehensive benefit of computing power reduces the cost of using computing power, making the computing power in the region more inclusive and benefiting more users.
The third is the reliability of diversified supply. Computing power integration means that different computing units, multiple architectures coexist, and multiple software and hardware are compatible. The new computing power infrastructure is open, diverse, and compatible, which is equivalent to “not putting eggs in one basket” and can reduce supply chain costs. Uncertain risks can make computing power more secure and reliable in the long run.
The fusion of computing power is equivalent to a martial arts master. Regardless of Huashan swordsmanship, Shaolin Kung Fu, Wudang Tai Chi, all kinds of “body skills” can be easily mastered, so it is natural to be able to face the challenges of any kind of computing tasks with ease.
Vertical and horizontal 2.0 stage: what kind of computing base is needed for the children of digital rivers and lakes
In the coming wave of digital economy, everyone has to become a son and daughter of the rivers and lakes and follow the trend. At this time, a universal computing power base is like the “God Assist” who is proficient in martial arts and body skills of all schools, and can provide full-precision and diverse Computing power allows users to cope with business changes and challenges more calmly.
Dismantling the problem of “computing partial science”, where will the intelligent computing center go next?
The National Industrial Information Security Development Research Center pointed out in the “Report” that in the 1.0 stage of rapid expansion from 2017 to 2021, the Intelligent Computing Center mainly provides specialized computing power. Entering the 2.0 stage from 2022, it is necessary to use the heterogeneous overlap of accelerator chips such as CPU and GPU to realize the integrated supply of high-precision general-purpose computing power and low-precision special-purpose computing power.
It can be seen that the next step of the Smart Computing Center is to develop towards a general-purpose computing power base. To build such a universal computing power base, there are two basic requirements of “one horizontal and one vertical”:
Yiheng: Multiple computing power requires multiple architectures, and the intelligent computing center must be “horizontally” compatible.
In the 1.0 stage of the Intelligent Computing Center, a vertically integrated chimney-style solution is adopted to specifically meet different application loads such as high-performance computing, artificial intelligence, and big data computing, and the versatility and compatibility are relatively low. Diversified general-purpose computing power bases need to achieve horizontal compatibility of chip platforms with different architectures, algorithm models on different fields, and data sets, so that users can choose the most suitable computing power solution according to business scenarios and computing tasks.
Dismantling the problem of “computing partial science”, where will the intelligent computing center go next?
One vertical: General computing power requires software and hardware collaboration, and the intelligent computing center must be “vertically” coupled.
Chips, algorithms, models, applications and other elements of different technical routes need to connect the upstream and downstream of the industry chain to solve the problem of software and hardware compatibility. Chip manufacturers, mid-range manufacturers and software developers, through software optimization, architecture integration and software Hardware collaboration to improve the overall performance of computing.
Breaking the vertical integration model and moving towards “horizontal” compatibility with “vertical” coupling has become the key to the 2.0 era of intelligent computing centers, thereby supporting thousands of industries to gallop in the digital age.
Seeing the tricks: the future challenges of the Intelligent Computing Center
The 1.0 stage of intelligent computing centers is rapidly expanding. Statistics show that more than 30 cities in China are currently building or proposing to build intelligent computing centers. 80% of the scenarios and computing power resources in the future digital economy will be carried by the intelligent computing center. This is a huge opportunity, but the greater the opportunity, the greater the responsibility that should be assumed.
As mentioned earlier, in the 2.0 stage, the Intelligent Computing Center must move towards the goal of “horizontal” compatibility with “vertical” coupling. Meals must be eaten at one bite, and computing power integration must be done step by step. Every link must be done well. trick.
Specifically, the computing power fusion of the intelligent computing center must work from four core links, namely: computing power production, computing power aggregation, computing power scheduling, and computing power release.
Dismantling the problem of “computing partial science”, where will the intelligent computing center go next?
In terms of computing power production, it is necessary to have the supply capacity of diversified computing power. In order to realize the diversified integration of chips with different architectures, the architecture of the Intelligent Computing Center has changed from homogeneous computing to heterogeneous computing. Domestic and foreign technology companies such as Google, Intel, Ali, and Baidu are all conducting research and development on heterogeneous computing. In the “Intelligent Computing Center 2.0 Era Prospect Report”, Sugon’s 5A-level intelligent computing center has also become a demonstration example in the industry by virtue of full computing power precision coverage and diversified computing power supply
In terms of computing power aggregation, it is necessary to promote the integration of general computing power and special computing power. At present, the construction concepts of intelligent computing centers of various manufacturers in the industry are relatively similar, and they all emphasize integrated development.
In terms of computing power scheduling, how can the integrated computing power be flexibly and leanly used by all walks of life? This requires unified scheduling and operation of diverse computing power. From a macro point of view, under the impetus of the East-to-West Computation Project, the national computing power integration network system has been initially formed, which has established a foundation for the computing power scheduling of the Intelligent Computing Center. From a micro perspective, the builders/participants of the Intelligent Computing Center have also begun to build a computing power operation platform, using refined and intelligent means to improve the level of computing power operation and scheduling, such as the national integrated computing power service platform operated by Suguang Smart Computing wait.
In terms of computing power release, the essential purpose of computing power fusion is to enable thousands of industries to use computing power well and use computing power correctly. Therefore, how to more flexibly release diversified computing power to the digital industry and provide applications for computing power services is crucial. The top priority of the Intelligent Computing Center in the 2.0 stage.
where will the intelligent computing center go next?
Dismantling the problem of “computing partial science”, where will the intelligent computing center go next?
From production to use, and dismantling of tricks, Smart Computing Center 2.0 will be a public computing power infrastructure in the true sense, handing over surging internal power and adaptable agility to thousands of industries.
A brand-new digital vision has already sent heroic posts to knights from all walks of life. The vast industrial opportunities are waiting for everyone to fight for and win their own medals in this era.
Use diversified computing power for your own use and open up a diversified future. In the 2.0 stage of the Intelligent Computing Center, there will definitely be more legendary people and events.