AWS Unveils Next-Gen Trainium3 AI Chip and Launches Trainium2-Powered Cloud Instances

10 Sources

Share

Amazon Web Services announces its next-generation AI chip, Trainium3, promising 4x performance boost over Trainium2. The company also launches Trainium2-powered cloud instances for high-performance AI computing.

News article

AWS Unveils Trainium3: Next-Generation AI Accelerator

Amazon Web Services (AWS) has announced its next-generation AI accelerator, Trainium3, at the re:Invent conference. Set to launch in late 2025, Trainium3 promises significant advancements in AI computing capabilities

1

2

.

Key features of Trainium3 include:

  • Built on a 3nm process node, expected to be the first dedicated machine learning accelerator using this technology

    2

  • 4x higher performance than its predecessor, Trainium2

    1

    3

  • 40% improvement in efficiency compared to Trainium2

    2

Trainium2 Enters General Availability

While Trainium3 is on the horizon, AWS has made Trainium2-powered cloud instances generally available:

  • EC2 Trn2 instances feature 16 Trainium2 processors

    1

  • Deliver up to 20.8 FP8 PetaFLOPS of performance

    1

  • Offer 1.5 TB of HBM3 memory with a peak bandwidth of 46 TB/s

    1

  • Provide 30-40% better price-performance over current GPU-based instances

    4

Scaling Up: Trn2 UltraServers and Project Rainier

AWS is pushing the boundaries of AI computing with larger configurations:

  • EC2 Trn2 UltraServers: 64 interconnected Trainium2 chips offering 83.2 FP8 PetaFLOPS

    1

  • Project Rainier: A collaboration with Anthropic to build a massive EC2 UltraCluster using hundreds of thousands of Trainium2 processors

    1

    4

Performance Comparisons and Industry Impact

The introduction of Trainium2 and Trainium3 positions AWS as a strong competitor in the AI chip market:

  • A single Trainium2 chip offers 1.3 PetaFLOPS of FP8 performance, comparable to Nvidia's H100 (1.98 PetaFLOPS)

    1

  • The EC2 UltraCluster could potentially deliver around 130 FP8 ExaFLOPS, equivalent to about 32,768 Nvidia H100 processors

    1

Broader AI Infrastructure Developments

AWS is not solely relying on its custom silicon:

  • The company continues to support various GPU instances, including Nvidia's H200, L40S, and L4 accelerators

    4

  • Project Ceiba: A massive AI supercomputer using Nvidia's Grace-Blackwell Superchips, expected to produce 414 exaFLOPS of super low precision sparse FP4 compute

    4

Industry Implications and Future Outlook

The development of Trainium3 and the scaling of Trainium2 instances signify AWS's commitment to advancing AI computing capabilities:

  • These advancements could potentially accelerate the development and deployment of larger, more sophisticated AI models

    5

  • The improved efficiency and performance may lead to cost reductions and faster time-to-market for AI-driven applications

    4

  • AWS's partnership with Anthropic for Project Rainier demonstrates the practical application of these technologies in pushing the boundaries of AI model training

    4

    5

As the AI chip race intensifies, AWS's innovations in custom silicon and cloud infrastructure are poised to play a crucial role in shaping the future of AI computing and applications.

TheOutpost.ai

Your Daily Dose of Curated AI News

Don’t drown in AI news. We cut through the noise - filtering, ranking and summarizing the most important AI news, breakthroughs and research daily. Spend less time searching for the latest in AI and get straight to action.

© 2025 Triveous Technologies Private Limited
Instagram logo
LinkedIn logo