Amazon has unveiled its generative artificial intelligence (AI) model, 'Nova.' Its subsidiary, Amazon Web Services (AWS), also announced plans to launch a new AI semiconductor, 'Trainium3,' next year.
Andy Jassy, Amazon's Chief Executive Officer (CEO), announced the release of the generative AI foundation model Nova during his keynote speech at AWS re:Invent 2024 on the 3rd (local time). He said, "I am pleased to share the launch of Amazon Nova, a new cutting-edge technology for everyday life."
Amazon Nova was launched in Micro and Multimodal versions such as Light and Pro. A higher-performance Premier version is scheduled for release in the first quarter of next year.
Nova Micro is a text-only model that provides the shortest response time at a low cost. Nova Light accepts image, video, and text inputs and outputs text, offering a fast and ultra-low-cost model. Nova Pro is a high-performance multimodal model with an optimal combination of accuracy, speed, and cost for various tasks. Nova Premier is designed for complex reasoning tasks and customized model derivation.
The Amazon Nova models also include creative content generation models such as Canvas and Reel. Nova Canvas features various editing functions like inpainting, outpainting, and background removal, allowing precise control over style and content. It is a state-of-the-art image generation model that produces studio-quality images.
Nova Reel is Amazon's first video generation model. It creates short videos from text prompts and images, controls visual style and speed, and can produce professional-quality video content for marketing, advertising, and entertainment.
CEO Jassy explained that the Nova Speech-to-Speech model and Any-to-Any model are planned for release in the first quarter and mid-year of next year, respectively. The Speech-to-Speech model allows users to ask questions by voice and receive answers by voice. Any-to-Any is a versatile model that can generate videos from text input or images from video input, supporting all combinations.
Additionally, Matt Garman, AWS CEO, announced plans to launch the Trainium3 semiconductor used for AI model training. Amazon stated that the Trainium3-based UltraServer can deliver performance four times better than the Trn2 Ultra server.
On this day, Garman CEO revealed for the first time that they are collaborating long-term with Apple regarding products such as Trainium3. Benoit Dupin, Apple's Senior Director of Machine Learning and AI, said, "Using AWS's Graviton3 (server chip), we improved efficiency by over 40%, and with the inference chip Inferentia2, we more than doubled efficiency."
Garman CEO also mentioned that to address the AI model drawback of hallucination, they introduced automated reasoning checks. He further unveiled the new high-performance managed relational database service 'Aurora DSQL' and storage service 'Amazon S3' for the first time and added that the P6 instance will be launched early next year.
Meanwhile, AWS stated that it is supporting the domestic AI startup Twelve Labs in enhancing its video search AI model capabilities. Twelve Labs has developed an AI foundation model that matches natural language with elements such as actions, objects, and background sounds occurring within video content, supporting the creation of applications capable of video search, scene classification, summarization, and video clip chapter segmentation.
Twelve Labs uses 'Amazon SageMaker HyperPod,' an AI model development and deployment service, to train foundation models that can simultaneously understand various data formats such as video, image, audio, and text. Training tasks are distributed across multiple AWS 'computing instances' operating in parallel, enabling uninterrupted foundation model training for weeks or months.
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
