Skip to main content
AWS logo beside a rack of Trainium3 UltraServers, with a Trainium4 chip diagram and a 6-FP4 performance boost label.

Editorial illustration for AWS Unveils Trainium3 UltraServers, Teases Trainium4 with 6x Performance Boost

AWS Trainium3 Servers Supercharge AI Computing Power

AWS unveils Trainium3 UltraServers, previews Trainium4 with 6× FP4 boost

Updated: 3 min read

AWS announced its Trainium3 UltraServers today. More importantly, the company has now deployed over one million of its custom Trainium accelerators. That scale is the foundation for its real play: using NVLink Fusion inside its MGX racks to directly wire those chips to NVIDIA GPUs and its own Graviton processors.

This builds an integrated system. And the teased Trainium4 chip, promising at least six times the FP4 processing punch, shows where this arsenal is headed next.

AWS also revealed early details of Trainium4, expected to deliver at least 6x the processing performance in FP4, along with higher FP8 performance and memory bandwidth. The next-generation chip will support NVIDIA NVLink Fusion interconnects to operate alongside NVIDIA GPUs and AWS Graviton processors in MGX racks. AWS has already deployed more than 1 million Trainium chips to date.

The company says the latest performance improvements translate to faster training and lower inference latency. In internal tests using OpenAI's GPT-OSS open-weight model, Trn3 UltraServers delivered three times higher throughput per chip and four times faster response times compared to Trn2 UltraServers.

Take that internal GPT-OSS benchmark. Triple the throughput. Response times four times faster.

Those numbers prove the performance leap from Trainium2 to Trainium3. But the strategic shift is the hardware integration. NVLink Fusion physically binds AWS silicon to NVIDIA’s dominant GPUs.

With a million chips already in circulation, this is beyond testing. It is AWS building a complete, scaled hardware stack—a concrete alternative to relying solely on external suppliers.

Common Questions Answered

What performance improvements can users expect from AWS's Trainium3 UltraServers?

AWS's Trainium3 UltraServers represent a significant leap forward in machine learning infrastructure with enhanced computational capabilities. The new servers are designed to improve AI workload processing, demonstrating AWS's commitment to custom silicon development for advanced AI computing.

How does AWS plan to integrate Trainium4 with existing GPU technologies?

Trainium4 will support NVIDIA NVLink Fusion interconnects, allowing seamless operation alongside NVIDIA GPUs and AWS Graviton processors in MGX racks. This integration strategy enables a more flexible and powerful multi-processor ecosystem for handling complex AI workloads.

What performance boost is AWS promising with the upcoming Trainium4 chip?

AWS is teasing Trainium4 with a promised 6x performance improvement in FP4 processing, along with enhanced FP8 performance and increased memory bandwidth. These advancements are expected to translate into faster AI training times and lower inference latency for users.

LIVE03:21OpenAI's Miles Wang in Talks for USD 2B AI Drug Discovery Startup