The world's largest AI supercomputer, next-generation AI chips and models, training Apple's AI... Amazon re:Invent global conference is here!
Amazon has launched the next-generation foundational model Amazon Nova and announced a partnership with Anthropic to develop the world's largest AI supercomputer. The next-generation, more advanced 3-nanometer training chip Trainium 3 will be released by the end of next year, with performance four times that of existing chips. In addition, Apple is conducting preliminary tests on the Trainium 2 chip to assist in training Apple Intelligence
As the world's largest cloud computing provider, Amazon is at a disadvantage in its recent competition with Microsoft and Google. To regain its momentum, Amazon is accelerating its efforts to build an advantage in the field of artificial intelligence computing.
On Wednesday local time, Amazon held its Re:Invent conference in Las Vegas and announced a series of major plans and products for AWS (Amazon Web Services), including:
Collaboration with OpenAI's rival Anthropic to create the world's most powerful AI supercomputer; preparation to collaborate with Apple to train Apple Intelligence; and the launch of the next-generation, more advanced 3-nanometer training chip Trainium 3 by the end of next year. Additionally, Amazon introduced the next-generation foundational model Amazon Nova.
Partnering with Anthropic to Build an AI Supercomputer
Matt Garman, CEO of Amazon Web Services, stated at the conference that the company is creating an AI training platform called UltraServers—comprising 64 Trainium 2 chips—to scale generative AI workloads.
This cloud service platform will be used for an ambitious plan—an AI supercomputer. Amazon announced at the conference that it will collaborate with Anthropic to develop the AI supercomputer UltraCluster, codenamed Rainer.
Amazon stated, it is expected that this computer will be equipped with hundreds of thousands of Amazon's latest AI training chips Trainium 2, and once completed, it will be five times larger than Anthropic's current most powerful model cluster, making it the world's largest AI supercomputer.
In the face of strong competition from Microsoft and OpenAI, Amazon has chosen to "huddle together for warmth" with Anthropic, having invested $8 billion in Anthropic this year.
More Powerful Next-Generation Chips to Challenge NVIDIA's Dominance?
Garman stated at the Re:Invent conference, Trainium 2 will be widely available in the Trn2 UltraServer clusters specifically designed for training cutting-edge AI.
Many companies have begun to use Amazon's cloud to build and train custom AI models, often in conjunction with NVIDIA's GPUs. However, Garman noted that the new AWS clusters are 30% to 40% cheaper than those using NVIDIA GPUs.
Garman stated that Amazon's goal is to reduce the cost of AI.
Today, there is really only one choice in GPUs, and that is NVIDIA. We believe customers will appreciate having multiple options.
We hope Trainium can carve out a good niche, and I think it will be an excellent choice for many workloads—not all workloads.
At the conference, Amazon also announced its next-generation, more advanced 3-nanometer training chip Trainium 3, which is said to have four times the performance of existing chips and is scheduled to be launched to customers by the end of 2025. Patrick Moorhead, CEO and Chief Analyst of Moore Insight & Strategy, stated that the data from Trainium 3 is "quite impressive," and its performance seems to have significantly improved due to enhancements in chip interconnects. Interconnects are crucial for developing ultra-large AI models, enabling fast data transfer between chips, and AWS appears to have optimized this in its latest design.
Moorhead also mentioned that NVIDIA may still be the dominant player in AI training for the foreseeable future, but it will face increasingly fierce competition in the coming years, as Amazon's innovations "indicate that NVIDIA is not the only player in the training space."
A New Chapter in Collaboration with Apple?
According to previous media reports, Apple is considering using the Trainium 2 chip for Apple Intelligence training. Garman stated at the conference that Apple is one of its clients and has adopted several AWS chips, such as Amazon Graviton and Inferentia, to support services including Siri.
Apple's Senior Director of AI and Machine Learning, Benoit Dupin, then took the stage at the conference in Las Vegas. He stated, Apple has collaborated with AWS in "almost every phase" of the AI and machine learning lifecycle.
AWS has been able to keep pace, and we have been customers for over a decade.
Dupin indicated that Apple is conducting preliminary tests on the Trainium 2 chip to assist in training Apple Intelligence.
Next-Generation Generative AI Models
At the conference, Amazon also launched a new generation of foundational models—Amazon Nova.
Amazon stated that AWS customers will be able to use generative AI applications supported by Amazon Nova to "understand videos, charts, and documents, or generate videos and other multimedia content," with a range of models available at different costs.
Among them, Amazon Nova Micro, Lite, and Pro support over 200 languages. Amazon Nova Micro supports a context length of 128K input tokens, while Amazon Nova Lite and Pro can reach a context length of 300K tokens, processing video lengths of up to 30 minutes.
“Amazon Nova Micro, Amazon Nova Lite, and Amazon Nova Pro are at least 75% cheaper than the best-performing models in their respective intelligence categories in Amazon Bedrock.”
Additionally, Amazon introduced two brand new models—Amazon Nova Canvas for generating high-quality images and Amazon Nova Reel for generating high-quality videos Rohit Prasad, Senior Vice President of General Artificial Intelligence at Amazon, revealed that approximately 1,000 generative AI applications are underway within Amazon.
AI Agent Administrators, Reasoning Inspectors...
Amazon has also launched a series of AI tools on the AWS platform to help businesses leverage and master generative AI technology.
For example, a new cloud tool called Bedrock Agents, announced on Wednesday, can be used to create and manage AI agents that can automatically perform useful tasks such as customer support, order processing, and analysis. It includes a main agent that manages the AI subordinate team, provides reports on how it operates, and assists in adjustments.
Garman said:
You can basically create an agent that makes you the boss of all other agents.
In addition, Amazon will also launch a system that enhances the capabilities of smaller models using larger models, as well as a tool to verify whether chatbot outputs are correct (to combat AI hallucinations)