AWS Continues to Invest in the Core of its Infrastructure

AWS is holding its annual AWS re:Invent conference in Las Vegas this week. As has become traditional, the provider is announcing a whole host of new features covering all the areas in which it is working to provide cloud services, although this time we will focus on the main building blocks that the company has been offering practically since its inception: compute, storage and databases.

These are the building blocks around which practically all the other services revolve to make up an increasingly extensive offering that allows any type of organisation, be it an SME or a large corporation, to focus on its business while consuming the latest technology through the public cloud model provided by AWS.

AWS CEO Matt Garman himself used the start of his need to talk about ‘building blocks’ as part of the provider’s success: ‘Almost any application can be broken down into individual components or core services, which we call “building blocks”. The idea was that if you had services that were the best in the world at doing a specific task, and if they could be easily combined in new and unique ways, then people could create very interesting things. This is how a new model for consuming technology and building companies was established.

‘This concept of ‘building blocks’ has been central to the creation of AWS services over the past 18 years and, perhaps more importantly, all of our software development practices. Today, we have hundreds of AWS services that users combine in unique and innovative ways,’ added Gartman, who proceeded to break down the first new features of AWS re:Invent 2024.

EC2, stronger in AI

On the compute side, AWS announced the general availability of EC2 Trn2 instances, based on the new Trainium2 chip. In addition, it introduced Trn2 UltraServers systems, designed to train and deploy advanced artificial intelligence models, such as large language models (LLMs) and foundational models (FMs), with superior performance and reduced costs. The company has also provided details of the upcoming Trainium3 chip, as explained below.

* EC2 Trn2 instances with Trainium2: EC2 Trn2 instances offer 30-40% better price-performance than current EC2 P5 GPU instances. Each instance combines 16 Trainium2 chips, providing up to 20.8 petaflops of compute capacity. They are designed to train and deploy models with billions of parameters, while maintaining high cost and time efficiency.

* Trn2 UltraServers: These new solutions combine four Trn2 servers in a NeuronLink interconnected structure, totalling 64 Trainium2 chips with a maximum capacity of 83.2 petaflops. The UltraServers offer unprecedented performance and scalability, facilitating the training of larger models in less time. AWS is developing an UltraCluster with Anthropic, called Project Rainier, which will include hundreds of thousands of Trainium2 chips to train the world’s most advanced models.

* Trainium3: The next generation of AI training chips will be built on 3 nanometre technology and promises 4x the performance of Trainium2. The first instances with Trainium3 are expected to be available by the end of 2025.

* Neuron SDK: AWS provides the Neuron SDK development kit to optimise the use of Trainium2 and help professionals build AI-based services. This kit includes compilers and tools compatible with popular frameworks such as PyTorch and JAX. With support for over 100,000 models in Hugging Face, it aims to facilitate the adoption of Trainium for developers.

New features coming to Amazon S3 storage

AWS also announced new features for Amazon S3, highlighting innovations that significantly improve data analytics and metadata management. These updates position S3 as the first cloud storage system with fully managed support for Apache Iceberg tables, an open source table format widely used in tabular data analytics, as we’ll see below:

* Amazon S3 Tables: AWS has introduced a new type of bucket optimised for tabular data stored as Apache Iceberg tables. This functionality improves query performance by up to three times and increases transactions per second (TPS) by up to ten times compared to traditional S3 buckets. S3 tables automate critical tasks such as data compaction and snapshot management, making it easier for organisations to handle large volumes of data without the need for dedicated maintenance teams. Companies like Genesys plan to use S3 Tables to simplify complex workflows and improve data analysis efficiency.

* Amazon S3 Metadata: This new feature enables the automatic generation of queryable metadata in near real-time, eliminating the need to build external systems to capture and manage metadata. Users can query this metadata using SQL, making it easy to search, organise and prepare data for analytics or artificial intelligence applications. Companies such as Roche plan to leverage S3 Metadata to manage large volumes of unstructured data in generative AI projects, improving the identification and use of relevant datasets.

Key benefits:

Optimised Performance: S3 Tables offers advanced transaction capabilities and fast queries for data lakes, with full support for analytics tools such as Amazon Athena and Apache Spark.
Simplified Management: Enhancements automate table and metadata management, reducing operational complexity.
AI Innovation: These updates are essential to power AI-based applications and real-time analytics.

SQL and NoSQL databases

The third ‘building block’ we discuss in these lines refers to databases, another key element for AWS. On this occasion, the provider has announced new capabilities for Amazon Aurora and Amazon DynamoDB, designed to improve performance, availability and consistency in globally distributed applications. These updates are focused on making it easier to manage demanding workloads, offering high availability, strong cross-region consistency and low latency for both SQL and NoSQL databases.

* Amazon Aurora DSQL: This is a new serverless, distributed SQL database that offers near-unlimited scalability, greater consistency across multiple regions, and 99.999% availability. Aurora DSQL offers up to four times faster read and write performance compared to other distributed SQL databases. Its active-active architecture and automatic failover ensure high availability, allowing applications to read and write anywhere in Aurora DSQL, with real-time synchronisation between regions. The solution is compatible with PostgreSQL, making it easy for developers to adopt.
According to the company, Aurora DSQL addresses historical distributed database challenges, such as synchronising transactions across regions with minimal latency. It uses Amazon Time Sync Service to ensure microsecond accuracy in global data synchronisation, crucial for critical applications such as those in the financial sector.

* Amazon DynamoDB Global Tables with greater consistency across regions: DynamoDB, a fully managed NoSQL database, now includes support for strong consistency in its global tables. This capability ensures that distributed applications always read the latest data without needing to modify code. DynamoDB global tables already offered 99.999% availability and unlimited scalability; now, with strong consistency, it becomes a more robust solution for critical applications that require real-time accuracy.

Key Benefits:

• High Performance: Aurora DSQL and DynamoDB offer significant improvements in read and write speed, as well as strong consistency across regions.
• Scalability and Availability: Both solutions eliminate the need to manage infrastructure, allowing organisations to scale without limits.
• Technical Innovation: The combination of precise synchronisation and active-active architecture offers new opportunities to build resilient global applications.

Over the next few days, we will be expanding our coverage of AWS re:Invent 2024 to bring readers the rest of the new features showcased by the public cloud provider in Las Vegas.