Amazon AWS today announced that AI training based on its internal team's development of thechip Wide availability of Trn2 instances of Trainium2 and the launch of the Trn2 UltraServer large-scale AI training system, as well as the release of the next generation of the more advanced 3nm process Trainium3 chip.
A single Trn2 instance consists of 16 Trainium2 chips interconnected by ultra-high-speed, high-bandwidth, low-latency NeuronLink, providing 20.8 petaflops of peak power for training and deploying B-parameter-sized models.
Amazon claims that Trn2 instances compare favorably to the current generation of GPU-based EC2 P5e and P5en instancesImproved value for money 30-40%.
The larger Trn2 UltraServer aggregates four Trn2 servers with NeuronLink, containing a total of 64 Trainium2 chips, and the peak arithmetic power further scales linearly to 83.2 petaflops, while the larger Trn2 UltraServer aggregates four Trn2 servers with NeuronLink, containing a total of 64 Trainium2 chips.Capable of meeting the training and deployment needs of the world's largest models today.
Amazon is also working with its AI modeling venture Anthropic to build a giant EC2 UltraCluster compute cluster called Project Rainier, which contains a large number of Trn2 UltraServers.Hundreds of thousands of Trainium2 chips in total..
1AI has learned that upon completion of the clusterPromising to be the largest publicly available AI computing cluster to dateThis is more than five times the overall power Anthropic currently requires to train the state-of-the-art Claude model.
Amazon AWS also announced its next-generation Trainium3 AI training chip, the first AWS chip product to use a 3nm process. Amazon saidTrainium3-based UltraServer delivers up to 4x the performance of Trn2 UltraServerThe first Trainium3-based instances are expected to be available by the end of 2025.