摘要：Along with the new Trainium2 EC2 instances, AWS introduced its next-generation Trainium3 AI processors, designed to enhance perfor
TMTPOST — Amazon announced the public launch of its new AWS Trainium2-powered Amazon Elastic Compute Cloud (EC2) instances at the AWS re:Invent conference. These GPU-based instances promise a significant 30-40% improvement in price-to-performance ratio compared to current offerings, according to a statement from Amazon.
Along with the new Trainium2 EC2 instances, AWS introduced its next-generation Trainium3 AI processors, designed to enhance performance and efficiency in future AI deployments. AWS also revealed Trn2 UltraServers, which are intended to support large-scale AI workloads.
The Trainium2 instances, which feature 16 Trainium2 processors, deliver 20.8 peak petaflops of computational power, making them ideal for training and deploying generative artificial intelligence projects, especially large language models. To further enhance these capabilities, the newly launched Trn2 UltraServers connect 64 Trainium2 processors via the ultra-fast NeuronLink interface, boosting processing capacity to up to 83.2 peak petaflops. This configuration aims to accelerate AI model training times and enable real-time inference for models with trillions of parameters.
Amazon also unveiled a suite of artificial intelligence models, marking its strongest push yet to compete with tech giants in the rapidly expanding generative AI market.
Rohit Prasad, Amazon’s senior vice president overseeing the company’s AI initiatives, explained, “Inside Amazon, we have about 1,000 Gen AI applications in motion, and we've had a bird's-eye view of what application builders are still grappling with. Our new Amazon Nova models are intended to help with these challenges.”
The Nova family includes six AI models designed to handle tasks ranging from text creation to video generation. Amazon claims the models are at least 75% more affordable than comparable offerings available on AWS servers, with improved processing speeds.
The lineup includes Nova Micro for fast text processing, Nova Lite for basic multimedia tasks, and Nova Premiere, scheduled for release in early 2025, for more complex reasoning. The models support 200 languages and offer customization options using customers' proprietary data, a feature Amazon hopes will attract enterprises seeking specialized AI applications.
Rohit Prasad, Amazon's head of artificial general intelligence, emphasized that the company would compete with rivals on both price and capabilities, showcasing the faster speeds of its new models. He noted that "if I have something better to offer, then customers will come and use it," underscoring Amazon's potential to lead in AI as the field continues to evolve.
A particularly exciting development is the Nova Reel software, which allows users to generate six-second videos from a single image or text prompt, ideal for product displays on Amazon’s website. The tool will soon be expanded to support videos up to two minutes long. The entertainment industry has shown keen interest in these AI tools to enhance filmmaking capabilities, although concerns about potential copyright infringement remain.
Amazon also introduced Canvas, an image-generation tool powered by short text prompts, and committed to using watermarking technology to ensure responsible usage and prevent the spread of harmful content.
Other offerings include advancements to accelerate the processing and analysis of text, and Amazon’s upcoming AI model that will support text, images, speech, and video input, allowing for the generation of any of these outputs.
来源：钛媒体