AWS Neuron 2.30.0 Adds Trainium3 Capabilities and New NKI Kernels
AWS Neuron 2.30.0 is now generally available, introducing new hardware capabilities for AWS Trainium3 and 22 new NKI Library kernels, enhancing model porting and optimization. This release offers features like scalar engine instructions, FP8 support, and expanded Agentic Development skills, benefiting ML developers working with Trainium and Inferentia instances. The update also includes a Neuron DRA Driver for Kubernetes and performance improvements for the Graph Compiler and Runtime, with availability across all regions supporting Neuron instances.
- →AWS Neuron 2.30.0 enhances Trainium3 hardware support with NKI 0.4.0
- →NKI Library gains 22 new kernels for various ML workloads
- →Neuron Agentic Development skills enhanced for model porting and validation
- →Neuron DRA Driver for Kubernetes introduced for topology-aware scheduling
- →Neuron Graph Compiler and Runtime see performance improvements
Features (4) ›
- AWS Neuron 2.30.0 enhances Trainium3 hardware support with NKI 0.4.0
The release includes the activate2 Scalar Engine instruction for Trn3, OCP FP8 input support for matrix multiplication, and bytes-aware tile-size constants to simplify kernel development for AWS Trainium3.
- NKI Library gains 22 new kernels for various ML workloads
The NKI Library now features 3 new core kernels for segmented attention, KV-parallel prefill, and FP8 quantization, along with 19 experimental kernels for advanced techniques like context parallelism and state-space models.
- Neuron Agentic Development skills enhanced for model porting and validation
New skills include neuron-framework-autoport for end-to-end HuggingFace model porting to NxD Inference and neuron-framework-equivalence for numerical validation of ported models. Both are now included by default in Neuron DLAMIs and Deep Learning Containers.
- Neuron DRA Driver for Kubernetes introduced for topology-aware scheduling
The Neuron DRA Driver enables dynamic resource allocation in Kubernetes, allowing for topology-aware scheduling of Trainium accelerators and Elastic Fabric Adapter (EFA) interfaces.
Enhancements (1) ›
- Neuron Graph Compiler and Runtime see performance improvements
The Neuron Graph Compiler delivers significant compile-time improvements, while the Neuron Runtime now enables zero-copy host-device transfers by default.
Notes (1) ›
- AWS Neuron 2.30.0 is available in all regions supporting Neuron instances
AWS Neuron is available in all AWS Regions where Amazon EC2 Trn1, Trn2, Inf2, and Inf1 instances are supported. PyTorch reference implementations are available for 29 kernels.
https://aws.amazon.com/about-aws/whats-new/2026/05/aws-announce-neuron-2-30-0
Related releases
- CloudWatch adds managed Prometheus collectors AWS What's New ·
- Amazon EC2 C7i Instances Expand to New Regions AWS What's New ·
- EC2 C7i-flex instances now available in Europe (Milan) region AWS What's New ·
- AWS CodeDeploy Expands to Five New Regions AWS What's New ·
- EC2 Auto Scaling Instance Refresh integrated with CloudFormation AWS What's New ·
- EC2 Dedicated Hosts support host resource groups without self-managed licenses AWS What's New ·