In addition to announcing the expansion of the application scope of self-made processors, more servers were introduced to meet the needs of artificial intelligence execution.Customized designAt the re:Invent 2024 event, AWS not only announced that it will introduce NVIDIA's code-named "Blackwell" GPU design in its new Amazon EC2 P6 Instances servers, but also revealed that it will officially launch the Trainium3 processor designed with a 3nm process next year, and will significantly upgrade the AWS cloud service infrastructure to give users more innovative "choices."
Propose innovations based on AWS user behavior
Matt Garman, who has taken over the AWS CEO position from current Amazon CEO Andy Jassy, stated that while this is his first time participating in re:Invent as AWS CEO, he emphasized that he has witnessed the changes AWS has driven in cloud application development over the years, as well as its continued innovation based on user behavior and its ability to help startups accelerate their technological growth.
By stacking technologies in line with AWS's "Building Blocks" development model, Matt Garman stated that hundreds of AWS application technologies are already being used to help more businesses accelerate their growth through the cloud platform. These include Pinterest, which aggregates popular trends and ideas, and EvolutionaryScale, which uses the ESM3 model to study protein structure. Even Apple, a long-term AWS partner, uses the AWS cloud platform to support services such as iCloud, including the recently launched "Apple Intelligence" service and even its search service.
To support the cloud platform needs of many enterprises and startups, AWS continues to update its service platform, including by introducing the Graviton series of self-made Arm architecture processors, which not only greatly improves the operating performance of its servers, but also reduces overall power consumption and carbon emissions.
Emphasizes continued cooperation with NVIDIA, but also continues to expand its own infrastructure
At the same time, AWS stated that it continues to maintain in-depth cooperation with NVIDIA. After Google and Microsoft announced the introduction of NVIDIA's code-named "Blackwell" GPU design in their servers, AWS also announced at re:Invent 2024 that it will use this GPU design in its new Amazon EC2 P6 Instance server, and it is expected to start deploying applications in the first quarter of 2025.
To meet the needs of AI inference, AWS not only uses its newly launched Trainium2 custom processor to improve computing performance, but also stacks four Amazon EC4 Trainium2 Instance servers equipped with this processor to form an even more powerful super server, the "Amazon EC2 Trainium2 UltraServer." The company also stated that Adobe, AI startup Poolside, data platform service Databricks, and Qualcomm are all using Trainium2 processors to train their AI models.
In response to the demand for artificial intelligence model inference execution, AWS announced that it will begin deploying its Trainium3 self-made processor built with a 3nm process next year, claiming to reduce energy consumption by 40% and increase execution performance by 2 times.
Upgrading various infrastructure designs with artificial intelligence computing
To meet the needs of artificial intelligence computing, AWS has upgraded many of its infrastructures, including providing appropriate storage environments for different computing architectures. It claims to help users save more than $3 billion in storage costs through the Amazon S40 Intelligent-Tiering storage category.
In addition, AWS announced support for the Apache Iceberg open source storage table format and the addition of the Amazon S3 Tables table system, which can increase storage speed by 3 times. Through the Amazon S3 Metadata management system, metadata comparison can be used to find the correct stored data more quickly, even if the data changes in the future.
To meet the data processing needs of providing services for global use, AWS also proposes to use satellites to correct errors in data processing time caused by different time zones, so that all data can correctly clarify their respective storage and modification times, and thus make different processing decisions. It also uses Amazon DynamoDB global tables to use a fully managed, serverless, multi-regional, multi-purpose database with 99.999% availability.
Other parts use Amazon Bedrock Model Distillation to quickly build more efficient AI models, and Amazon Bedrock Automated Reasoning Checks to reduce human errors, confirm that the artificial intelligence model is operating correctly, and provide correction and adjustment suggestions. Amazon Bedrock multi-agent collaboration can quickly deploy assistant services that can handle complex, multi-step processes.
The Amazon Q Developer feature previously provided to developers has also been integrated with the software development platform GitLab Duo. It also provides the ability to migrate services used in Microsoft .NET environments to Linux and increase execution speed by 4 times. It can also convert processes originally in VMware environments into cloud-native types and reduce the integration time of application services on mainframes from several years to several months.
Amazon Q Business features now include the ability to analyze unstructured data, enabling integration of more data. Developers can also use APIs to automate complex processes and design applications using Amazon Q Business features. The new version of Amazon SageMaker will analyze data through tools such as Unified Studio, Zero-ETL for Applications, and Amazon SageMaker Lakehouse, enabling data analysis results to accelerate AI operational efficiency.
