AWS and Cerebras Unveil Disaggregated Inference to Shatter AI Speed Records

Amazon Web Services (AWS) and Cerebras Systems have announced a pioneering collaboration aimed at setting a new industry standard for AI inference speed. By introducing a "disaggregated inference" architecture available exclusively through Amazon Bedrock, the partnership tackles the computational bottlenecks of generative AI. The solution separates the AI workload into two stages: AWS Trainium chips will handle prompt processing (prefill), while the Cerebras CS-3 system—boasting unprecedented memory bandwidth—will tackle token generation (decode). Connected via Amazon’s Elastic Fabric Adapter (EFA), this specialized hardware division promises to deliver AI inference an order of magnitude faster than current market solutions, benefiting data-heavy enterprise applications like agentic coding.

Read the original article on Business Wire here