Trainium2
AWS second-generation AI training accelerator chip
Trainium2 is the second-generation machine learning accelerator from Amazon Web Services. It features eight NeuronCore-V3 units and supports large-scale clusters for training generative AI models with improved efficiency.
- Contains eight NeuronCore-V3 cores per chip
- Supports clusters of over 100,000 chips
- Provides 96 GiB of device memory
- Successor to the first-generation Trainium
What it is
Trainium2 is a purpose-built integrated circuit designed by Amazon Web Services for machine learning workloads. It serves as the second generation of the Trainium family, succeeding the original Trainium chip. The processor targets artificial intelligence training and inference tasks, particularly for large language models and generative AI applications. Each chip integrates eight NeuronCore-V3 processing units to handle complex mathematical operations required by modern neural networks.
How it came about
Amazon Web Services announced Trainium2 alongside its Graviton4 central processing unit. The launch aimed to provide cloud customers with an alternative to third-party graphics processors for AI development. The company stated that this design offers significant performance gains over its predecessor while improving energy efficiency. This release marked a step in AWS strategy to reduce reliance on external silicon suppliers for high-demand artificial intelligence infrastructure.
Inside the chip
The architecture relies on eight NeuronCore-V3 cores per chip. These cores support various data formats including FP8, FP16, BF16, and TF32. A key feature is Logical NeuronCore Configuration, which allows combining resources from multiple physical cores into one logical unit. Sixteen CC-Cores manage communication between chips. The design includes NeuronLink-v3 for high-speed interconnects, enabling memory pooling and efficient scaling across many devices in a cluster.
How it was made
Specific manufacturing details such as the foundry or process node size are not provided in the available documentation. The chip contains 96 gibibytes of device memory with high bandwidth capabilities. It supports direct memory access speeds of 3.5 terabytes per second. The interconnect bandwidth reaches 1.28 terabytes per second per chip, facilitating rapid data exchange in multi-chip configurations.
Where it went
These accelerators power Amazon Elastic Compute Cloud instances dedicated to AI model training. They are deployed in clusters of sixteen chips within standard instances. Larger Trn2 UltraServers combine sixty-four processors for massive workloads. Companies like Anthropic use these systems to train Claude models. Other platforms such as Databricks and Hugging Face have integrated support to help developers build and scale artificial intelligence solutions more cost-effectively.
Versions and family
Trainium2 is the successor to the first-generation Trainium chip. AWS has also announced Trainium3 as the next iteration, expected in late 2025. The current generation supports scaling up to over one hundred thousand chips in a single cluster. This scalability distinguishes it from smaller deployments, allowing for the training of models with trillions of parameters.
Why it matters
The introduction of Trainium2 represents a move by cloud providers to develop custom silicon for specific workloads. By offering these instances, AWS aims to lower the cost and energy consumption associated with training large artificial intelligence models. The ability to scale to hundreds of thousands of chips supports the growing demand for generative AI capabilities without relying solely on general-purpose graphics hardware.
Questions people ask about Trainium2
What is the primary purpose of Trainium2?
Trainium2 is designed specifically for machine learning tasks. It accelerates the training and inference of artificial intelligence models, including large language models and generative AI applications, within Amazon Web Services cloud infrastructure.
How many cores does a single Trainium2 chip have?
Each Trainium2 chip contains eight NeuronCore-V3 cores. These cores can be configured logically to combine compute and memory resources, allowing for flexible handling of different AI workload requirements.
Can Trainium2 chips be connected together?
Yes, they use NeuronLink-v3 for chip-to-chip interconnects. This allows them to be scaled into large clusters, with AWS supporting configurations of up to over one hundred thousand chips for massive AI training jobs.
Good reading and sources
There is no English Wikipedia article on Trainium2 yet, so the facts on this page come from the sources listed above: awsdocs-neuron.readthedocs-hosted.com/en/latest/general/arch/neuron-ha, servethehome.com/aws-graviton4-is-an-even-bigger-arm-server-processor-, finance.yahoo.com/news/aws-launches-trainium2-instances-improved-19234. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.