Amazon
youtube
AWS and Cerebras announced a collaboration to set a new standard for AI inference speed and performance in the Cloud, which will be available through Amazon Bedrock.
The solution combines AWS Trainium3-powered servers, Cerebras CS-3 systems, and Elastic Fabric Adapter (EFA) networking. The Trainium3 + CS-3 solution enables “inference disaggregation,” a technique which separates AI inference into two stages: prompt processing, or “prefill,” and output generation, or “decode.” These two stages have profoundly different computational characteristics. Prefill is natively parallel, computationally intensive, and requires moderate memory bandwidth. Decode, on the other hand, is inherently serial, computationally light, and memory bandwidth intensive. Decode typically represents the majority of inference time in these scenarios because each output token must be generated sequentially. Together, we're leveraging the fastest system for each stage of inference. Trainium3 handles compute-intensive prefill, and Cerebras's wafer-scale CS-3 handles memory-intensive decode. Each stage runs on the hardware it excels at. The result is the fastest inference in Amazon Bedrock. Later this year, AWS will also offer leading open-source LLMs and Amazon Nova using Cerebras hardware.
Learn more:
Subscribe to AWS:
Create a free AWS account:
Try AWS for free:
Connect with an expert:
Explore more:
Next steps:
Explore on AWS in Analyst Research:
Discover, deploy, and manage software that
|
2026/03/13
youtube
