The latest pattern of China's AI platform is released! Baidu's comprehensive score is the first, the second echelon is fiercely competitive, and the large model accelerates the evolution of cloud vendors
The trend of large models has brought hundreds of millions of "small shocks" to everything related to it.
An AI/ML platform is one of them.
It is closely related to the trend of large models, and can directly reflect the level of AI technology R&D reserves of major cloud vendors, as well as the insight and understanding ability of the latest trends.
Who is stronger? It is being talked about by the industry.
Under the drastic changes in the technological wind, AI/ML platforms also have new evaluation standards.
Forrester's latest "First Chinese Artificial Intelligence/Machine Learning Platform Report" released by the international authority provides timely reference.
Forrester Research is an independent technology and market research firm that publishes thematic reports that are highly recognized in China and around the world.
The Forrester Wave, published biennially, is Forrester's most influential report type.
The report surveyed 14 mainstream cloud vendors in the domestic market, including Baidu Intelligent Cloud, Alibaba Cloud, Huawei Cloud, Tencent Cloud, etc., and evaluated them from three aspects: product capabilities, strategic planning, and market performance.
Based on a comprehensive evaluation of 25 segments, Forrester divides the 14 leading vendors into four quadrants: Leaders, Top Performers, Contenders, and Challengers.
Let's take a look at the specific highlights.
What new standards do new trends bring
Let's start with the report's core conclusions.
In this quadrant chart, the strategic level is the horizontal axis, the product capability is the vertical axis, and it is divided into four quadrants: leaders, excellent performers, competitors and challengers, and it also reflects the market performance of each company.
The distribution of the quadrants is as follows:
Leaders (2), Top Performers (5), Competitors (4), Challengers (3).
In the first echelon are Baidu Intelligent Cloud and Alibaba Cloud. Among them, Baidu Intelligent Cloud performed brilliantly and won the first place in the comprehensive score.
The second echelon is the most competitive, with 5 vendors in a very compact position in the quadrant.
The above conclusion is the result of 25 evaluations conducted by Forrester.
In addition to sorting out the current competitive landscape of AI/ML platforms in the Chinese market, Forrester has further proposed a new standard reference for evaluating AI/ML platforms.
The following three points are the most critical:
Comprehensive toolchain
Easy-to-use accelerator
Model Ops at scale
Why?
Forrester believes that under the influence of generative AI and large model trends, AI applications are becoming more important to promote productivity and accelerate business innovation.
In today's Chinese market, companies are in dire need of an AI/ML platform that can solve complex problems within their own business environment.
In order to meet the market demand, the three aspects mentioned above are indispensable.
First, there must be a toolchain that provides data management, model training, and AI application development capabilities. **
This is also the core of the AI/ML platform.
Forrester proposes that the platform side should not only pay attention to the key tools in model construction, training and evaluation, but also pay attention to the tools required for AI application development.
For example, AI frameworks and notebooks for professionals; Low-code, visual tools for business people.
In addition to this, data management tools also have a significant impact on model building.
Second, it provides an easy-to-use accelerator for the industry. **
Forrester points out that most companies in the domestic market lack engineers who understand both AI algorithms and business knowledge, which makes it difficult for them to tailor algorithms to their business needs.
At present, they are embracing the "pose" of the large model trend, either using the large model for fine-tuning, or doing prompt engineering.
Therefore, acceleration tools that can accelerate AI model construction and application development are very important. Visualization tools, low-code development, and more can further accelerate innovation.
Third, accelerate the implementation of large models through large-scale model operations. **
Model Ops includes model deployment, monitoring, updating, and automation, which can solve problems such as model drift, performance degradation, security maintenance, and model updates, and provide A/B testing, automatic tuning, and model retraining.
As enterprises embrace the trend of large models, AI/ML platforms with large-scale model operations can better help enterprises develop, deploy, and manage AI models, further promote enterprise digital transformation, reduce costs and increase efficiency.
In summary, for an AI/ML platform to become a leader in the new trend, it needs to provide easy-to-use tools, meet industry needs, and accelerate the adoption of AI by enterprises.
And what capabilities are more specific to be possessed, we have to analyze from the current leaders.
How do you make it to the Leaders quadrant? **
In this Forrester report, the performance of Baidu Intelligent Cloud is eye-catching.
It is the only cloud vendor to enter the "Leader" quadrant, and it also won the first place in the overall score and the first place in the 9 subdivision scores.
In terms of product capabilities, Baidu Intelligent Cloud is at the leading level in four subdivisions: data, training, predictive reasoning, and application.
At the same time, in terms of strategic dimension and market size, Baidu also won the first place in many of these projects.
Forrester describes Baidu Intelligent Cloud as "one of the pioneers of China's basic model":
Baidu Intelligent Cloud embeds the ERNIE family of foundation models into a portfolio with a solid product roadmap, and its active ecosystem around PaddlePaddle is an effective way to engage AI developers to co-innovate.
Its specific capabilities come from Baidu's AI platform, and its products include BML, EasyDL and Baidu's intelligent cloud Qianfan large model platform.
Taking Baidu's intelligent cloud AI platform as an example, we can analyze more specifically what capabilities the current AI/ML platform should have if it wants to occupy a leading position in the market.
According to the dimensions of the Forrester report, specific capabilities can be divided into five aspects: data, training, predictive reasoning, application, and architecture. These are also the five core elements in the development and application of AI models.
**Let's start with the data. **
In the data processing part, Baidu's AI platform can process both structured and unstructured data.
More than 65 types of data visualizations are supported, including pie charts, heat maps, scatter plots, maps, and more. It supports 10+ kinds of filter components, and users only need to make simple configuration and drag-and-drop to achieve real-time data monitoring and auxiliary decision-making.
At the same time, it also supports 30+ data format annotations, improving the efficiency of the annotation link with the most concentrated manpower in the modeling process.
It is worth mentioning that the Baidu AI platform providesactive learning annotationcapability, the system can directly analyze the pattern of pictures from the dataset, automatically filter out the most critical pictures, and prompt limited annotation.
For example, if there are 10,000 pictures that need to be labeled, the system will put the characteristic pictures in front and the repetitive ones behind, so that only the first 3,000 pictures will be marked, and the last 7,000 pictures can be automatically marked.
According to reports, this method can save 70-90% of manpower for enterprise users on average. The proportion of human labor in "artificial intelligence" has been drastically reduced.
In addition, in terms of feature engineering, Baidu's AI platform integrates the professional-level feature database management capability, providing functions such as feature addition, deletion, modification and query, feature production, feature sharing, feature version management, and data verification.
Different forms of data in batch and streaming are supported for prediction services, which can ensure that the characteristics of the model are consistent during training and final prediction, which is directly related to the accuracy of the model.
The above capabilities are reflected in specific numbers, and the Forrester report gives Baidu's AI platform a score of 5 out of 5 (out of a score) for its data capabilities, significantly ahead of other vendors.
**The second is model training. **
This is one of the most obvious aspects of the latest trend in terms of market demand, which is aimed at not only professional developers, but also business people who are not specialized in AI algorithms, so it requires that the AI/ML platform that provides capabilities is easy to use and flexible.
Refer to the practice of Baidu's AI platform.
On the one hand, it attaches importance to "breadth" and supports the modeling and training of a variety of data, such as images, videos, text, and speech.
The modeling method is also very flexible, supporting a variety of modeling methods such as notebook/WebIDE development, drag-and-drop visual development, script parameter tuning, custom jobs, etc., which can complete the customized development of high-precision models for people with different professional levels.
On the other side is "depth". With the support of its own paddle algorithm team, Baidu's AI platform has made in-depth optimization of a large number of scene operators. Including image classification, object detection, text classification, sequence annotation, etc.
For example, PP YOLO, which is deeply optimized based on the Paddle operator, has surpassed the benchmark YOLO V3 in the field of object detection.
Ease of use is also a big focus. On the Baidu AI platform, zero-code modeling and visual modeling can be realized. In the former, users only need to upload data and select a type to start modeling. The latter can assemble a modeling process by dragging and dropping the components and setting the parameters of each part.
In addition, the amount of AI computing is increasing at least 10 times per year, and the ability to adjust task resources in deep learning training has become particularly important. Baidu's AI platform supports multi-machine and multi-card distributed training, and provides a variety of types of computing resources.
In addition, Baidu itself has rich experience in training super-large models, and can integrate its own capabilities in visual large models, generative AI, etc. For example, automatic hyperparameter search, automatic processing of unbalanced data, and ultra-large-scale pre-training can be performed.
As a result, on the Baidu AI platform, you can also see a lot of development tools to improve programming efficiency.
The third dimension of competence is reasoning. **
With the development of large models, the inference market will further expand and even grow exponentially, which poses great challenges to AI/ML platforms.
From the perspective of Baidu's AI platform, they mainly focus on development efficiency, performance optimization, flexibility, and extensiveness.
Its inference module, Model Serve, supports 16 AI frameworks, including the most common Paddle, Tensor Flow, and PyTorch, as well as Matlab/R for scientific computing and Xg boost for machine learning.
In terms of performance optimization, an asynchronous inference scheduler is directly abstracted on the scheduling layer to achieve heterogeneous inference workers, improving the performance of the entire server and GPU utilization by more than 1 times.
At the same time, it supports automatic batch processing, classifies tasks of different lengths, and compiles tasks of similar sizes into the same batch to make full use of heterogeneous resources, which can improve the efficiency by 70% on the basis of asynchronous decoupling.
The fourth area to focus on is the application. **
The application in the report mainly examines the application efficiency of each platform.
That is, how to use existing resources to quickly turn data into business productivity.
Baidu's AI platform can provide full lifecycle management capabilities for the AI development process, from data collection and cleaning, to model development and training, model management, to cloud and offline inference service management.
It is worth mentioning that the Baidu AI platform is the first platform in China to reach the flagship level of the MLOps standard of the Academy of Information and Communications Technology.
At present, the capabilities of Baidu's AI platform have been exported to industries such as finance, energy, and transportation. Serving Shanghai Pudong Development Bank, Bank of Beijing, as well as State Grid, China Southern Power Grid, etc.
In 2022, the number of public cloud paying users of Baidu's intelligent cloud AI platform will increase by 49%, the number of privatized customers will increase by 32%, and the number of developers will increase by 1.228 million, with a growth rate of about 40%; The repurchase rate has increased year after year, and the repurchase rate in key industries has reached 50%.
Finally, in terms of architecture, the architecture design of Baidu's AI platform received a full score in the Forrester score.
If it can be summed up in one sentence, Baidu's AI platform has reached the level of "leader", that is, it has done it:
There are many algorithms, tools, fast operation, good results, and it also saves servers and manpower, while ensuring security and ease of use.
Through the analysis of the overall capabilities of Baidu's AI platform, it is not difficult to find that many of these tools and ideas are in line with the new needs of the current large model trend.
In fact, in the context of drastic changes in the direction of technology, it is not only the trend to adjust the existing architecture to adapt to the changes in demand, but also to propose new ways to deal with it.
In the era of large models, the new competition pattern of AI on the cloud has been preliminarily determined
So, with the impact of the wave of large models, what new changes have taken place in the market demand for AI/ML platforms?
In the past, many models of CV and NLP were called SOTA, but they were still more often used in non-core businesses in the industry. Now, with its amazing ability to subvert traditional workflows, large models are beginning to be recognized more and more, and are considered to be the key to breaking through the bottleneck of intelligence in various industries.
However, for cloud vendors, this does not mean that the era of large models is "starting from scratch" in the era of small models.
In fact, with the deepening of the application of large models, technical fields such as agents have attracted more and more attention. The core lies in the fact that the large model is based on its own capabilities and connects the small model with mature scheduling to solve the problem in practical applications, which is considered to be faster and more valuable in the production scenario.
Therefore, in the "new era" opened by large models, for the "leaders" of AI/ML platforms, the technical accumulation in the era of small models and the technological innovation in the era of large models are complementary and indispensable.
Baidu's AI platform in the "new era" handed over the answer sheet - Baidu Intelligent Cloud Qianfan Large Model Platform is an example.
As a one-stop enterprise-level large model platform, Baidu Intelligent Cloud Qianfan Platform is essentially the product of Baidu's deep accumulation in the chip layer, framework layer, model layer and application layer.
It is embodied in five aspects:
First, at the level of computing power, Baidu's intelligent cloud Qianfan platform can provide efficient and cost-effective heterogeneous computing services.
In the process of large model training, through the distributed parallel training strategy and microsecond-level interconnection capabilities, the acceleration ratio of 10,000-card scale cluster training on Baidu Qianfan platform can reach 95%. At the same time, the effective training time of the Vanka cluster can reach 96%, which greatly reduces the cost of computing power and time.
Second, at the model level, Baidu Qianfan platform has managed 44 mainstream large models at home and abroad, including Wenxin large model, Llama series, ChatGLM, etc., and supports users to quickly call APIs and directly obtain large model capabilities.
For third-party large models, Baidu Qianfan platform has also been optimized in a targeted manner, including Chinese enhancement, performance enhancement, context enhancement and so on.
Baidu revealed that the number of large model API calls on Baidu's Qianfan platform is continuing to rise at a high speed. At present, Baidu Qianfan platform has served more than 20,000 customers.
Third, for customers who want to carry out secondary development based on existing large models, Baidu Qianfan platform provides a full life cycle tool chain for the retraining, fine-tuning, evaluation and deployment of large models, as well as 41 high-quality datasets, which can realize rapid model optimization for specific business scenarios.
Fourth, at the application level, in response to the needs of enterprises to develop AI-native applications based on large models, Baidu Qianfan platform provides a series of capability components and frameworks.
For example, there are 226 built-in templates, so that developers can quickly improve the quality of large model answers even if they are not familiar with prompt engineering.
At the Baidu World Conference on October 17, Baidu Intelligent Cloud also released the "Baidu Qianfan AI Native Application Development Workbench". Specifically, this "workbench" consists of two parts: application components + application frameworks.
**Application component services are composed of two components: AI and basic cloud. **
Among them, the AI component, that is, the component-based encapsulation of large model capabilities, includes large language model components such as Q&A and Chain of Thought (CoT), as well as multimodal components such as Wensheng diagram and speech recognition.
The basic cloud components include traditional cloud services such as vector databases and object storage.
Application framework is oriented to specific scenario tasks, which can be understood as an effective combination of the above application components based on the capabilities of large models.
At present, Baidu Qianfan platform provides commonly used AI native application frameworks such as Retrieval Enhanced Generation (RAG) and Agent.
Among them, the RAG framework can combine the knowledge in the enterprise's proprietary domain with the large model Q&A ability to make more accurate answers to professional knowledge.
Based on this RAG framework, Sany Heavy Industry quickly realized the development and launch of the intelligent customer service application on the official website.
Shen Dou, executive vice president of Baidu Group and president of Baidu Intelligent Cloud Business Group, revealed that even if you need to process thousands of long-word documents, the cost of building such a "little assistant" is only a few hundred yuan; After that, the cost of each information for users is only a few cents.
Based on the agent framework, the large model can automatically disassemble the tasks given by humans, automatically plan and call various components to complete the tasks collaboratively, and at the same time provide self-feedback according to the task completion effect to improve its own capabilities.
At present, based on this agent framework, Zhongtian Iron and Steel has built an intelligent "enterprise scheduling center" to realize the automatic perception, decomposition and execution of task instructions.
For example, when it is found that the steel output is not up to standard, the large model can automatically call various resources and APIs managed by the platform to find out the reason for the non-compliance, adjust the production schedule in time and send an email to notify the dispatcher.
Finally, Baidu Qianfan also launched the "AI Native App Store", which connects the supply side and the demand side of AI native applications, and provides a gathering place for large-scale model business opportunities.
It is not difficult to see that, on the one hand, the rapid launch of Baidu's intelligent cloud Qianfan large model platform benefits from the development of Baidu's own large model technology; On the other hand, the product capabilities accumulated by Baidu's AI platform over the years, as well as its rich practical experience in the industry, have made Baidu's Qianfan platform take the lead in playing an effective role on the application side.
According to IDC data, China's AI public cloud service market will show a positive growth of 80.6% in 2022, with the overall market size reaching 7.97 billion yuan.
IDC analysis believes that the implementation of generative AI and large models is currently in its infancy, and these capabilities can be updated and iterated more quickly on the public cloud, which will bring significant benefits to AI public cloud services in the short term.
Gartner also points out that generative AI is driven by large models, which puts forward the requirements for a robust and highly scalable computing infrastructure. "The cloud provides the perfect solution and platform, and the key players in the generative AI race must be the top cloud vendors."
Combined with the latest report from Forrester, it can be seen that for cloud vendors, AI cloud services have become the new focus of competition.
And how to measure competitiveness, now the evaluation criteria are gradually clear.
In the final analysis, there are two core aspects:
First, from the perspective of developers and enterprise users, it is whether the ability of the AI cloud service platform can truly cost-effectively solve the practical problems faced by complex businesses, as well as the shortage of professional talents in the process of intelligent upgrading, especially under the wave of large models.
Second, from the perspective of technology trends, it is more closely integrated with large models.
The layout of Baidu's AI platform can be regarded as a reference answer given by leading AI cloud vendors in the latest changes in the competitive landscape.
As for the results? In more landing cases, you can see the real chapter.
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
The latest pattern of China's AI platform is released! Baidu's comprehensive score is the first, the second echelon is fiercely competitive, and the large model accelerates the evolution of cloud vendors
Source: Qubits
The trend of large models has brought hundreds of millions of "small shocks" to everything related to it.
An AI/ML platform is one of them.
It is closely related to the trend of large models, and can directly reflect the level of AI technology R&D reserves of major cloud vendors, as well as the insight and understanding ability of the latest trends.
Who is stronger? It is being talked about by the industry.
Forrester's latest "First Chinese Artificial Intelligence/Machine Learning Platform Report" released by the international authority provides timely reference.
Forrester Research is an independent technology and market research firm that publishes thematic reports that are highly recognized in China and around the world.
The Forrester Wave, published biennially, is Forrester's most influential report type.
Based on a comprehensive evaluation of 25 segments, Forrester divides the 14 leading vendors into four quadrants: Leaders, Top Performers, Contenders, and Challengers.
Let's take a look at the specific highlights.
What new standards do new trends bring
Let's start with the report's core conclusions.
In this quadrant chart, the strategic level is the horizontal axis, the product capability is the vertical axis, and it is divided into four quadrants: leaders, excellent performers, competitors and challengers, and it also reflects the market performance of each company.
Leaders (2), Top Performers (5), Competitors (4), Challengers (3).
In the first echelon are Baidu Intelligent Cloud and Alibaba Cloud. Among them, Baidu Intelligent Cloud performed brilliantly and won the first place in the comprehensive score.
The second echelon is the most competitive, with 5 vendors in a very compact position in the quadrant.
The above conclusion is the result of 25 evaluations conducted by Forrester.
In addition to sorting out the current competitive landscape of AI/ML platforms in the Chinese market, Forrester has further proposed a new standard reference for evaluating AI/ML platforms.
The following three points are the most critical:
Why?
Forrester believes that under the influence of generative AI and large model trends, AI applications are becoming more important to promote productivity and accelerate business innovation.
In today's Chinese market, companies are in dire need of an AI/ML platform that can solve complex problems within their own business environment.
In order to meet the market demand, the three aspects mentioned above are indispensable.
First, there must be a toolchain that provides data management, model training, and AI application development capabilities. **
This is also the core of the AI/ML platform.
Forrester proposes that the platform side should not only pay attention to the key tools in model construction, training and evaluation, but also pay attention to the tools required for AI application development.
For example, AI frameworks and notebooks for professionals; Low-code, visual tools for business people.
In addition to this, data management tools also have a significant impact on model building.
Forrester points out that most companies in the domestic market lack engineers who understand both AI algorithms and business knowledge, which makes it difficult for them to tailor algorithms to their business needs.
At present, they are embracing the "pose" of the large model trend, either using the large model for fine-tuning, or doing prompt engineering.
Therefore, acceleration tools that can accelerate AI model construction and application development are very important. Visualization tools, low-code development, and more can further accelerate innovation.
Model Ops includes model deployment, monitoring, updating, and automation, which can solve problems such as model drift, performance degradation, security maintenance, and model updates, and provide A/B testing, automatic tuning, and model retraining.
As enterprises embrace the trend of large models, AI/ML platforms with large-scale model operations can better help enterprises develop, deploy, and manage AI models, further promote enterprise digital transformation, reduce costs and increase efficiency.
And what capabilities are more specific to be possessed, we have to analyze from the current leaders.
How do you make it to the Leaders quadrant? **
In this Forrester report, the performance of Baidu Intelligent Cloud is eye-catching.
It is the only cloud vendor to enter the "Leader" quadrant, and it also won the first place in the overall score and the first place in the 9 subdivision scores.
In terms of product capabilities, Baidu Intelligent Cloud is at the leading level in four subdivisions: data, training, predictive reasoning, and application.
At the same time, in terms of strategic dimension and market size, Baidu also won the first place in many of these projects.
Its specific capabilities come from Baidu's AI platform, and its products include BML, EasyDL and Baidu's intelligent cloud Qianfan large model platform.
Taking Baidu's intelligent cloud AI platform as an example, we can analyze more specifically what capabilities the current AI/ML platform should have if it wants to occupy a leading position in the market.
According to the dimensions of the Forrester report, specific capabilities can be divided into five aspects: data, training, predictive reasoning, application, and architecture. These are also the five core elements in the development and application of AI models.
**Let's start with the data. **
In the data processing part, Baidu's AI platform can process both structured and unstructured data.
More than 65 types of data visualizations are supported, including pie charts, heat maps, scatter plots, maps, and more. It supports 10+ kinds of filter components, and users only need to make simple configuration and drag-and-drop to achieve real-time data monitoring and auxiliary decision-making.
At the same time, it also supports 30+ data format annotations, improving the efficiency of the annotation link with the most concentrated manpower in the modeling process.
For example, if there are 10,000 pictures that need to be labeled, the system will put the characteristic pictures in front and the repetitive ones behind, so that only the first 3,000 pictures will be marked, and the last 7,000 pictures can be automatically marked.
According to reports, this method can save 70-90% of manpower for enterprise users on average. The proportion of human labor in "artificial intelligence" has been drastically reduced.
In addition, in terms of feature engineering, Baidu's AI platform integrates the professional-level feature database management capability, providing functions such as feature addition, deletion, modification and query, feature production, feature sharing, feature version management, and data verification.
Different forms of data in batch and streaming are supported for prediction services, which can ensure that the characteristics of the model are consistent during training and final prediction, which is directly related to the accuracy of the model.
The above capabilities are reflected in specific numbers, and the Forrester report gives Baidu's AI platform a score of 5 out of 5 (out of a score) for its data capabilities, significantly ahead of other vendors.
This is one of the most obvious aspects of the latest trend in terms of market demand, which is aimed at not only professional developers, but also business people who are not specialized in AI algorithms, so it requires that the AI/ML platform that provides capabilities is easy to use and flexible.
Refer to the practice of Baidu's AI platform.
On the one hand, it attaches importance to "breadth" and supports the modeling and training of a variety of data, such as images, videos, text, and speech.
The modeling method is also very flexible, supporting a variety of modeling methods such as notebook/WebIDE development, drag-and-drop visual development, script parameter tuning, custom jobs, etc., which can complete the customized development of high-precision models for people with different professional levels.
On the other side is "depth". With the support of its own paddle algorithm team, Baidu's AI platform has made in-depth optimization of a large number of scene operators. Including image classification, object detection, text classification, sequence annotation, etc.
For example, PP YOLO, which is deeply optimized based on the Paddle operator, has surpassed the benchmark YOLO V3 in the field of object detection.
In addition, the amount of AI computing is increasing at least 10 times per year, and the ability to adjust task resources in deep learning training has become particularly important. Baidu's AI platform supports multi-machine and multi-card distributed training, and provides a variety of types of computing resources.
In addition, Baidu itself has rich experience in training super-large models, and can integrate its own capabilities in visual large models, generative AI, etc. For example, automatic hyperparameter search, automatic processing of unbalanced data, and ultra-large-scale pre-training can be performed.
As a result, on the Baidu AI platform, you can also see a lot of development tools to improve programming efficiency.
The third dimension of competence is reasoning. **
With the development of large models, the inference market will further expand and even grow exponentially, which poses great challenges to AI/ML platforms.
From the perspective of Baidu's AI platform, they mainly focus on development efficiency, performance optimization, flexibility, and extensiveness.
Its inference module, Model Serve, supports 16 AI frameworks, including the most common Paddle, Tensor Flow, and PyTorch, as well as Matlab/R for scientific computing and Xg boost for machine learning.
In terms of performance optimization, an asynchronous inference scheduler is directly abstracted on the scheduling layer to achieve heterogeneous inference workers, improving the performance of the entire server and GPU utilization by more than 1 times.
At the same time, it supports automatic batch processing, classifies tasks of different lengths, and compiles tasks of similar sizes into the same batch to make full use of heterogeneous resources, which can improve the efficiency by 70% on the basis of asynchronous decoupling.
The application in the report mainly examines the application efficiency of each platform.
That is, how to use existing resources to quickly turn data into business productivity.
Baidu's AI platform can provide full lifecycle management capabilities for the AI development process, from data collection and cleaning, to model development and training, model management, to cloud and offline inference service management.
At present, the capabilities of Baidu's AI platform have been exported to industries such as finance, energy, and transportation. Serving Shanghai Pudong Development Bank, Bank of Beijing, as well as State Grid, China Southern Power Grid, etc.
In 2022, the number of public cloud paying users of Baidu's intelligent cloud AI platform will increase by 49%, the number of privatized customers will increase by 32%, and the number of developers will increase by 1.228 million, with a growth rate of about 40%; The repurchase rate has increased year after year, and the repurchase rate in key industries has reached 50%.
Finally, in terms of architecture, the architecture design of Baidu's AI platform received a full score in the Forrester score.
Through the analysis of the overall capabilities of Baidu's AI platform, it is not difficult to find that many of these tools and ideas are in line with the new needs of the current large model trend.
In fact, in the context of drastic changes in the direction of technology, it is not only the trend to adjust the existing architecture to adapt to the changes in demand, but also to propose new ways to deal with it.
In the era of large models, the new competition pattern of AI on the cloud has been preliminarily determined
So, with the impact of the wave of large models, what new changes have taken place in the market demand for AI/ML platforms?
In the past, many models of CV and NLP were called SOTA, but they were still more often used in non-core businesses in the industry. Now, with its amazing ability to subvert traditional workflows, large models are beginning to be recognized more and more, and are considered to be the key to breaking through the bottleneck of intelligence in various industries.
However, for cloud vendors, this does not mean that the era of large models is "starting from scratch" in the era of small models.
In fact, with the deepening of the application of large models, technical fields such as agents have attracted more and more attention. The core lies in the fact that the large model is based on its own capabilities and connects the small model with mature scheduling to solve the problem in practical applications, which is considered to be faster and more valuable in the production scenario.
Therefore, in the "new era" opened by large models, for the "leaders" of AI/ML platforms, the technical accumulation in the era of small models and the technological innovation in the era of large models are complementary and indispensable.
As a one-stop enterprise-level large model platform, Baidu Intelligent Cloud Qianfan Platform is essentially the product of Baidu's deep accumulation in the chip layer, framework layer, model layer and application layer.
It is embodied in five aspects:
First, at the level of computing power, Baidu's intelligent cloud Qianfan platform can provide efficient and cost-effective heterogeneous computing services.
In the process of large model training, through the distributed parallel training strategy and microsecond-level interconnection capabilities, the acceleration ratio of 10,000-card scale cluster training on Baidu Qianfan platform can reach 95%. At the same time, the effective training time of the Vanka cluster can reach 96%, which greatly reduces the cost of computing power and time.
For third-party large models, Baidu Qianfan platform has also been optimized in a targeted manner, including Chinese enhancement, performance enhancement, context enhancement and so on.
Baidu revealed that the number of large model API calls on Baidu's Qianfan platform is continuing to rise at a high speed. At present, Baidu Qianfan platform has served more than 20,000 customers.
Third, for customers who want to carry out secondary development based on existing large models, Baidu Qianfan platform provides a full life cycle tool chain for the retraining, fine-tuning, evaluation and deployment of large models, as well as 41 high-quality datasets, which can realize rapid model optimization for specific business scenarios.
Fourth, at the application level, in response to the needs of enterprises to develop AI-native applications based on large models, Baidu Qianfan platform provides a series of capability components and frameworks.
For example, there are 226 built-in templates, so that developers can quickly improve the quality of large model answers even if they are not familiar with prompt engineering.
At the Baidu World Conference on October 17, Baidu Intelligent Cloud also released the "Baidu Qianfan AI Native Application Development Workbench". Specifically, this "workbench" consists of two parts: application components + application frameworks.
Among them, the AI component, that is, the component-based encapsulation of large model capabilities, includes large language model components such as Q&A and Chain of Thought (CoT), as well as multimodal components such as Wensheng diagram and speech recognition.
The basic cloud components include traditional cloud services such as vector databases and object storage.
Application framework is oriented to specific scenario tasks, which can be understood as an effective combination of the above application components based on the capabilities of large models.
At present, Baidu Qianfan platform provides commonly used AI native application frameworks such as Retrieval Enhanced Generation (RAG) and Agent.
Among them, the RAG framework can combine the knowledge in the enterprise's proprietary domain with the large model Q&A ability to make more accurate answers to professional knowledge.
Based on this RAG framework, Sany Heavy Industry quickly realized the development and launch of the intelligent customer service application on the official website.
Based on the agent framework, the large model can automatically disassemble the tasks given by humans, automatically plan and call various components to complete the tasks collaboratively, and at the same time provide self-feedback according to the task completion effect to improve its own capabilities.
At present, based on this agent framework, Zhongtian Iron and Steel has built an intelligent "enterprise scheduling center" to realize the automatic perception, decomposition and execution of task instructions.
For example, when it is found that the steel output is not up to standard, the large model can automatically call various resources and APIs managed by the platform to find out the reason for the non-compliance, adjust the production schedule in time and send an email to notify the dispatcher.
IDC analysis believes that the implementation of generative AI and large models is currently in its infancy, and these capabilities can be updated and iterated more quickly on the public cloud, which will bring significant benefits to AI public cloud services in the short term.
Gartner also points out that generative AI is driven by large models, which puts forward the requirements for a robust and highly scalable computing infrastructure. "The cloud provides the perfect solution and platform, and the key players in the generative AI race must be the top cloud vendors."
Combined with the latest report from Forrester, it can be seen that for cloud vendors, AI cloud services have become the new focus of competition.
And how to measure competitiveness, now the evaluation criteria are gradually clear.
In the final analysis, there are two core aspects:
First, from the perspective of developers and enterprise users, it is whether the ability of the AI cloud service platform can truly cost-effectively solve the practical problems faced by complex businesses, as well as the shortage of professional talents in the process of intelligent upgrading, especially under the wave of large models.
Second, from the perspective of technology trends, it is more closely integrated with large models.
The layout of Baidu's AI platform can be regarded as a reference answer given by leading AI cloud vendors in the latest changes in the competitive landscape.
As for the results? In more landing cases, you can see the real chapter.