AWS Trainium

Amazon EC2 Trn2 instances, powered by AWS Trainium2 chips, are purpose-built for high-performance deep learning training of generative AI models, including large language models and diffusion models. They offer up to 50% cost-to-train savings over comparable Amazon EC2 instances. Trn2 instances support up to 16 Trainium2 accelerators, providing up to 3 petaflops of FP16/BF16 compute power and 512 GB of high-bandwidth memory. To facilitate efficient data and model parallelism, Trn2 instances feature NeuronLink, a high-speed, nonblocking interconnect, and support up to 1600 Gbps of second-generation Elastic Fabric Adapter (EFAv2) network bandwidth. They are deployed in EC2 UltraClusters, enabling scaling up to 30,000 Trainium2 chips interconnected with a nonblocking petabit-scale network, delivering 6 exaflops of compute performance. The AWS Neuron SDK integrates natively with popular machine learning frameworks like PyTorch and TensorFlow.

Amazon EC2 Trn3 UltraServers are AWS’s newest accelerated computing instances, powered by the in-house Trainium3 AI chips and engineered specifically for high-performance deep-learning training and inference workloads. These UltraServers are offered in two configurations, a “Gen1” with 64 Trainium3 chips and a “Gen2” with up to 144 Trainium3 chips per UltraServer. The Gen2 configuration delivers up to 362 petaFLOPS of dense MXFP8 compute, 20 TB of HBM memory, and a staggering 706 TB/s of aggregate memory bandwidth, making it one of the highest-throughput AI compute platforms available. Interconnects between chips are handled by a new “NeuronSwitch-v1” fabric to support all-to-all communication patterns, which are especially important for large models, mixture-of-experts architectures, or large-scale distributed training.