Amazon announces new 3nm AI chip 'Trainium3', 4 times faster than 'Trainium2' and up to 50% lower cost, and also announces 'Trainium4'
Amazon Web Services (AWS) has begun offering the latest generation of Trainium chips, the Trainium3 , through its UltraServer service. Trainium3 is the company's first AI chip manufactured on a 3nm process node, and offers various performance improvements over the previous generation.
Top announcements of AWS re:Invent 2025 | AWS News Blog
Trainium3 UltraServer delivers faster AI training at lower cost
https://www.aboutamazon.com/news/aws/trainium-3-ultraserver-faster-ai-training-lower-cost
AWS makes Trainium3 UltraServers generally available - DCD
https://www.datacenterdynamics.com/en/news/aws-makes-trainium3-ultraservers-generally-available/
The Trainium3 UltraServer is a system specially designed for next-generation AI workloads, capable of accommodating up to 144 Trainium3 chips. It delivers up to 4.4x the computing performance, 4x the energy efficiency, and 3.9x the memory bandwidth compared to the previous generation Trainium2 UltraServer. This accelerates model training, reducing the time required for training from months to weeks, and simultaneously responds to user inference requests, enabling companies to tackle AI projects that were previously unfeasible or costly.
In tests using OpenAI's open weight model 'GPT-OSS,' the throughput per chip was three times higher and response time was four times faster than that of the Trainium2 UltraServer.
Customers can connect thousands of UltraServers to up to one million Trainium3 servers, roughly 10 times the number of servers used in the previous generation. According to AWS, Anthropic, Karakuri, Metagenomics, Neto.ai, Ricoh, and Splashmusic are already using Trainium3 UltraServers. However, some critics on the social networking site Hacker News have
On the same day, AWS also announced ' AWS AI Factories ,' a project to transform customers' existing infrastructure into a high-performance AI environment. This project deploys dedicated AWS AI infrastructure in customers' data centers and operates it exclusively for them, providing secure, low-latency access to computing, storage, databases, and AI services. Customers can now use their existing data center space and power capacity to access AWS's AI infrastructure and services. Trainium3 has also been deployed in this AWS AI Factories.
AI Factories: AWS brings AI infrastructure directly to customer data centers
Additionally, Amazon has revealed that it is already working on its next-generation chip, Trainium 4. Compared to Trainium 3, Trainium 4 is expected to have at least six times the FP4 processing performance, three times the FP8 processing performance, and four times the memory bandwidth, resulting in a fundamental leap in performance and improved training speed.
Additionally, Amazon has announced the Nova 2 series of cost-effective AI models across inference, multimodal processing, conversational AI, code generation, and agent tasks, as well as Nova Forge , a service that enables 'open training' to make it easier to customize AI models.
Meet new Amazon Nova AI models that help build highly reliable AI agents
https://www.aboutamazon.com/news/aws/aws-agentic-ai-amazon-bedrock-nova-models
The Nova 2 is divided into four models: Nova 2 Lite, Nova 2 Pro, Nova 2 Sonic, and Nova 2 Omni, each boasting high performance and cost-effectiveness.
According to Artificial Analysis , an AI evaluation platform, Nova 2.0 Pro Preview's token usage is lower than that of its peers, while Nova 2.0 Lite and Omni are cheaper than most other inference models.
Nova Forge is a service that allows customers to independently train models from an early stage by providing access to individual pre-trained, intermediate, and post-trained Nova models, avoiding the risk of unintended consequences from continuing to train a model without access to the original training data and the costly barriers of building a model from scratch.
Additionally, Nova Act, a new AWS service for building and managing reliable AI agents for UI-based workflows, is now available, enabling customers to take advantage of superior performance for UI-based workflows such as updating data in a customer relationship management (CRM) system, testing website functionality, or submitting health insurance claims.
Amazon has announced the new AI agents: the Kiro Autonomous Agent , the AWS Security Agent , and the AWS DevOps Agent .
Amazon launches frontier AI agents that work autonomously like teammates
Kiro is a software development agent that orchestrates human work by reconstructing context when switching between tasks, manually reconciling changes between repositories, and connecting information scattered across tickets, pull requests, and chat threads. You can ask Kiro questions, describe tasks, and assign them to your backlog directly from GitHub.
AWS Security Agents are secure app enablers that proactively identify risks throughout development and quickly respond when issues arise. They incorporate deep security expertise, proactively review design documents, and scan pull requests against organizational security requirements and common vulnerabilities.
AWS DevOps Agents are agents that improve operational efficiency by isolating problems when an outage occurs, understanding system behavior, and accurately identifying root causes to reduce mean time to resolution. They shorten recovery time by providing recommendations for improving four key areas: observability, infrastructure optimization, deployment pipeline enhancements, and application resilience.
In addition, Amazon Bedrock AgentCore , a platform for developing AI agents, has been revamped to provide an environment for building and deploying agents safely on a large scale, diagnostics to understand how agents perform in real-world environments, and an environment for building accurate agents by setting clear definitions and limiting the scope of agent behavior.
Amazon launches AI Agents that stay within boundaries, track their performance, and get smarter over time
https://www.aboutamazon.com/news/aws-amazon-bedrock-agent-core-ai-agents
In addition to this, new features have been announced from various Amazon-related services.
Related Posts:
in Hardware, Software, Web Service, Posted by log1p_kr