AWS, Cerebras Pair Trainium With CS-3 in AI Inference Architecture

Amazon Web Services (AWS) is working with AI chipmaker Cerebras Systems to integrate AWS's in-house Trainium accelerators with Cerebras' CS-3 system. The collaboration targets speed and cost bottlenecks in large-language-model inference, reflecting cloud providers' efforts to reduce their reliance on a single supplier through heterogeneous chip architectures and improve the efficiency of generative AI services.

The companies have unveiled an “inference disaggregation” architecture that splits prompt processing and output generation into two stages optimized for Trainium and CS-3, respectively. They aim to improve inference performance by an order of magnitude, or about 10 times. To date, AWS and Cerebras have not disclosed the value of the partnership, a formal commercial launch date or benchmarked cost data.