Trainium3
Fourth-generation AI accelerator chip from AWS
Trainium3 is a machine learning chip designed by Amazon Web Services for AI training and inference. Launched in late 2025, it offers four times the compute performance of its predecessor and supports high-bandwidth memory pooling across devices.
- Delivers 2,517 sparse TFLOPS in FP16/BF16/TF32
- Features 144 GiB memory with 4.9 TB/sec bandwidth
- Provides 4X compute boost over Trainium2
- Supports Logical NeuronCore Configuration for resource pooling
- Launched generally available at re:Invent 2025
What it is
Trainium3 is the fourth generation of purpose-built machine learning chips from Amazon Web Services. It serves as an accelerator for both training artificial intelligence models and running inference tasks. The chip integrates eight NeuronCore-v4 processing units and sixteen communication cores. AWS launched the hardware at its re:Invent conference in November 2025. The company positions the device as a high-efficiency platform for enterprise cloud workloads, aiming to improve cost performance and reduce latency compared to earlier generations.
How it came about
AWS developed Trainium3 as the successor to Trainium2, which saw significant adoption among partners like Anthropic. The new chip was announced as generally available during the re:Invent 2025 event in Las Vegas. CEO Matt Garman highlighted the launch as a major innovation for the cloud provider. The design builds on the traction of the previous generation, which already handled more than half of the inference tokens for the Bedrock service. The development focuses on supporting open ecosystems and agentic AI tools.
Inside the chip
The architecture features eight NeuronCore-v4 cores per device. These cores support dynamic shapes and control flow through specific instruction set extensions. Users can program rounding modes, choosing between nearest even or stochastic methods. The chip includes deeply embedded GPSIMD engines for custom operators. Sixteen CC-Cores manage collective communication within and between servers. Logical NeuronCore Configuration allows combining resources from multiple physical cores into one logical unit. This design facilitates efficient scale-out training and memory pooling across different devices.
How it was made
Specific manufacturing details such as the process node or transistor count are not disclosed in the provided documentation. The chip delivers 144 GiB of device memory with a bandwidth of 4.9 TB per second. Direct memory access bandwidth also reaches 4.9 TB per second with inline computation capabilities. Device-to-device interconnect via NeuronLink-v4 provides 2.56 TB per second of bandwidth. These specifications enable high-speed data movement required for large-scale model training and inference operations.
Where it went
Trainium3 powers AI training and inference workloads in AWS data centres. It supports the Bedrock service, where it handles model serving for enterprise customers. The chip is designed to integrate with agent development tools like Kiro and AgentCore. Users can deploy models from various providers, including Google Gemini, on this hardware. The platform aims to help development teams scale their output by providing efficient compute resources for autonomous agents and software development tasks.
Versions and family
Trainium3 is the direct successor to Trainium2. AWS claims it provides a fourfold increase in compute performance over the previous generation. The new chip offers improved latency, throughput, and cost efficiency for serving models. While Trainium2 remains in the market with substantial deployment, Trainium3 represents the current flagship for AWS custom silicon.
Why it matters
AWS executives describe Trainium3 as the best inference platform globally due to its efficiency and low latency. The company expects the business around this chip to grow significantly, potentially matching the scale of its EC2 service. By offering better price-performance options, AWS aims to capture more enterprise market share. The hardware supports an open strategy that allows customers to mix frameworks and models from different vendors, distinguishing it from competitors who may prefer closed systems.
Questions people ask about Trainium3
What is the primary use case for Trainium3?
Trainium3 is designed for both machine learning model training and inference. It accelerates AI workloads in AWS data centres, supporting services like Bedrock and enabling efficient deployment of large language models and autonomous agents.
How does Trainium3 compare to its predecessor?
AWS states that Trainium3 offers a four times compute boost over Trainium2. It also provides better latency, higher throughput, and lower costs for serving models, making it more efficient for enterprise inference tasks.
Can Trainium3 run models from other companies?
Yes. AWS promotes an open strategy where customers can use models from other providers, such as Google Gemini, on Trainium3. The platform supports mixing various frameworks and tools within the AWS ecosystem.
Good reading and sources
There is no English Wikipedia article on Trainium3 yet, so the facts on this page come from the sources listed above: awsdocs-neuron.readthedocs-hosted.com/en/latest/general/arch/neuron-ha, crn.com/news/ai/2025/aws-tranium3-ai-is-the-best-inference-platform-in. 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.