Text | Intelligent relativity

Author | Chen Bocheng

Amazon and Anthropic, Microsoft and OpenAI, and Huawei Cloud's large-model hybrid cloud concept all show that the trend of deep integration of cloud computing and large-scale models has become a consensus in the industry.

At present, more and more companies are accelerating the deployment and application of large model technology on cloud computing platforms, which in turn promotes the iteration and upgrading of the cloud computing industry.

From the underlying infrastructure to the platform services at the middle layer to the scenario applications at the top level, cloud computing is undergoing significant changes. Taking Amazon Web Services as an example, as a global cloud computing giant, his cloud products are comprehensively paving the way for the deployment and application of large models.

1. At the bottom level, build an infrastructure represented by GPUs and self-developed chips for training basic models and running inference in the production environment.

Second, in the middle layer, Amazon Bedrock was launched, based on a fully managed service, allowing users to easily access carefully screened third-party brand models, such as AI21 Labs, Anthropic, Cohere, Meta, Mistral AI, Stability AI, etc., as well as Amazon's own brand model Amazon Titan.

Third, at the top layer, build out-of-the-box generative AI applications such as Amazon Q through basic models, allowing users to quickly get started with generative AI without any professional knowledge.

From this point of view, the development of generative AI is changing the industrial landscape of cloud computing itself, and the competition in the entire market has reached an unprecedented stage of change.

Generative AI, reimagining the cloud

Adhering to similar views and actions of Amazon Web Services, there are also local cloud vendors.

Hou Zhenyu, vice president of Baidu Group, previously proposed that the large model will drive the innovation of cloud computing and reshape the industrial pattern of cloud computing, driving the reconstruction of the underlying IT infrastructure and bringing about the transformation of the upper-level application development model.

HUAWEI CLOUD Stack 8.3, the industry's first large-scale hybrid cloud in China, combines hybrid cloud with large-scale models to provide a new direction for the development of cloud services.

The advancement of this capability, based on the technical development of large-model hybrid cloud, is essentially a specialization of application scenarios.

For example, the continuous application and innovation of cloud-edge collaboration technology on the hybrid cloud of large models aims to solve the real-time inference needs of edge real-time inference exposed by the gradual application of AI large models to industrial scenarios such as coal mine production, power inspection, and industrial quality inspection.

On the one hand, industrial scenarios have more requirements for AI applications than other scenarios, and the efficiency requirements are also higher. On the other hand, when the AI model is applied in industrial scenarios, its version iteration and function upgrade will inevitably enter the stage of learning while using and applying while upgrading.

Therefore, cloud-edge collaboration has become the key, and it is necessary to balance computing resources, optimize data flow, improve processing efficiency, and enhance service quality, supporting diverse and demanding AI application scenarios.

Based on the ModelArts AI development platform and the Pangu model, Huawei's hybrid cloud cloud-based collaboration solution provides a one-stop scenario-based model training workflow. Then, by collecting the original production sample data and the doubtful sample data generated during the model operation, the model is efficiently trained using workflows, and the model version is managed in a unified manner, effectively realizing the AI model learning while using, rapid iteration, continuous upgrading, and adapting to new working conditions and data changes.

Speaking of data, data storage is a problem for the training efficiency of AI large models. With the increase in the number of parameters of large AI models, the scale of training clusters is also expanding, and traditional storage is already difficult to cope with the needs of ultra-large-scale AI clusters for fast data reading, fast checkpoint storage, and fast fault recovery.

Faced with these specific requirements, HUAWEI CLOUD has to seek breakthroughs in storage architecture, based on the innovative three-layer architecture of OBS data lake, SFS Turbo high-performance parallel file system, and AI Turbo acceleration, to systematically meet the challenges of large model training scenarios.

In general, just to deal with the various scenarios of large AI models, cloud services need to be comprehensively innovated from the underlying infrastructure to the top-level applications, and corresponding solutions should be proposed to further promote the development of large AI models. In recent years, HUAWEI CLOUD has been systematically innovating to solve the key blockages in the application of large models, and the ten innovative technologies it released include enhanced AI networks, operator acceleration, unified data encoding, and diversified computing power scheduling in addition to cloud-edge collaboration and data storage.

In fact, aside from the concept of a large-model hybrid cloud proposed by HUAWEI CLOUD, the industry has reached a consensus on the combination of cloud and large-scale models, and is committed to providing various technical solutions to solve various training, inference, and application requirements of large-scale models on the cloud.

For example, JD Cloud has launched a complete set of tools for large models, including the infrastructure to support the application of large models, such as Yanxi AI development and computing platform, vector database, hybrid multi-cloud operating system Cloudship, high-performance storage platform Yunhai, software and hardware integrated virtualization engine Jinggang and other core products, which corresponds to promoting the industrialization development of large models on the basis of cloud.

The systematic breakthrough of large-model hybrid cloud

Amazon Web Services, Huawei Cloud, JD Cloud, Baidu Intelligent Cloud and many other vendors are committed to creating complete technical solutions for today's large-scale model era, covering a series of processes and services such as the bottom, middle, and top-layer, so that the large-scale model can be continuously deployed and applied on the cloud and release value.

The concept of large-scale hybrid cloud has brought this comprehensive solution of cloud vendors to a more systematic stage. The formation of this kind of systematization not only requires cloud vendors to focus on technology, but also to conduct extensive exploration for scenarios.

"For the government, what they are concerned about may not be to simply solve the problem of saving a customer service personnel and operation and maintenance personnel internally, but to drive the development of the entire industry through large models from the perspective of industrial layout." Wu Bingkun, founder and CEO of Zhongshu Information Technology, said in an interview with the media.

Based on the development trend of the cloud service industry, the systematic development of this industry actually needs to be realized with the help of the systematic upgrade of cloud technology, which is a manifestation of the big picture. In this regard, from the perspective of traditional multi-cloud strategies, the proposal of large-scale hybrid cloud can better show the overall pattern of future cloud services.

A multi-cloud strategy focuses on using the services of multiple cloud service providers to avoid vendor lock-in while optimizing costs or leveraging the strengths of each service provider. Although a large-scale hybrid cloud may involve multiple cloud environments, its core lies in building a highly integrated and optimized large-scale data processing and AI model operating environment, not simply to disperse service sources, but to achieve specific technical and business goals.

For example, HUAWEI CLOUD Stack's multi-cloud collaborative architecture allows industry models to be trained on the public cloud, fine-tuned in hybrid clouds with local enterprise data, and then inferred at the edge cloud to meet computing requirements in different scenarios.

The essence of this is not to decentralize the "cloud", but based on the native hybrid cloud capability, users can extend the large model from local to edge and public cloud, realize cross-cloud deployment in all scenarios, and achieve optimization of application efficiency, security performance and other results.

Therefore, in summary, a large model hybrid cloud is a hybrid cloud architecture optimized for specific domains, especially scenarios that need to process large-scale data and complex AI models, which integrates the elasticity of the public cloud with the security controls of the private cloud, as well as the possibility of multi-cloud services, to meet the special needs of high-performance computing and AI applications.

The technical systematization realized by this integration will achieve the systematic development of the industry in the next time, that is, "it is not simply to solve the problem of saving a customer service personnel and operation and maintenance personnel internally, but to drive the development of the entire industry through a large model from the perspective of industrial layout." ”

Therefore, based on such systematic development, the development of large model + hybrid cloud will form several significant trends in the future.

1. In terms of computing power scheduling, the training and inference of large models usually require a large amount of computing resources. As the size of the model grows, so does the need for computing power. Computing power scheduling and optimization technologies in hybrid cloud environments will continue to evolve to support more efficient large model training and inference.

2. In terms of cloud-edge collaboration, edge computing is becoming more and more important with the popularization of Internet of Things (IoT) devices. The hybrid cloud architecture will support tighter cloud-edge collaboration, enabling large models to perform real-time inference at the edge, reducing latency and improving responsiveness.

3. In terms of infrastructure, AI-Native storage and network technologies will continue to evolve to support more efficient model training and inference processes. For example, high-performance storage supports a multi-level caching mechanism to achieve second-level access to checkpoints and minute-level recovery from training failures.

Fourth, in terms of model application, enterprises can fine-tune pre-trained large models through local data in the hybrid cloud environment to meet the needs of specific business scenarios while keeping the data private.

5. In terms of business deployment, different industries (such as finance, healthcare, manufacturing, etc.) will use the big model on the hybrid cloud to solve specific business challenges and promote business innovation and process automation. At the same time, large models are easier to deploy at scale in a hybrid cloud environment, especially in widely distributed enterprises and industries, such as energy, transportation, manufacturing, and other fields.

6. In terms of ecological construction, the combination of large models and hybrid cloud will promote more ecological partners to join in and jointly develop solutions and services to expand the entire ecosystem. At the same time, with the increase of large-scale hybrid cloud applications, relevant standards and protocols will be gradually formulated and improved to improve the interoperability and compatibility between different systems.

Write at the end

Today, large models have hundreds of millions or even billions of parameters, which provides unprecedented computational scale and complexity for the development of generative AI. More parameters mean that the model can learn deeper and more refined data features, so as to generate higher quality and diverse content in multiple fields such as text generation, image synthesis, and audio creation, which well promotes the high-quality development of generative AI.

For the industry, such ability is a key to the future industrial transformation and upgrading. Cloud computing is also the underlying technology for industrial upgrading, and the collaboration with generative AI will achieve this goal in a more comprehensive and complete form. However, in this process, how the cloud will be combined with the large models behind generative AI will be a key question.

Generative AI is a "free ride", and cloud vendors such as Amazon Web Services, Huawei Cloud, JD Cloud, and Baidu Intelligent Cloud all want to take it, but it takes some effort to build it.

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com