AMD and Cerebras partner on ultra-low-latency, disaggregated AI inference platform combining Helios EPYC with Cerebras Wafer-Scale Engine.

AMD and Cerebras Systems have formed a technical partnership to build a disaggregated artificial intelligence inference system that addresses speed bottlenecks in advanced software. The combined platform merges AMD Helios rack scale infrastructure with the Cerebras Wafer Scale Engine and is scheduled to launch on Cerebras Cloud in the second half of the year.

Modern AI workloads have distinct computational requirements depending on the task. Simple text generation demands massive data capacity, while real-time agents and coding assistants require instant feedback. The partnership optimizes for both by dividing the work between two specialized chips: AMD Helios handles initial prompt processing and large data contexts, while the Cerebras Wafer Scale Engine manages the memory-intensive task of generating individual words. This cooperative division of labor is expected to deliver 5x higher efficiency, measured in tokens per second per watt.

During the Advancing AI conference, AMD CEO Lisa Su underscored the industry's need for more adaptable computing architectures. Andrew Feldman, Cerebras's CEO, similarly highlighted the immediate market demand for faster processing capabilities.

The speed of text generation has become critical in software development, robotics, and scientific research, where sluggish response times degrade user experience. To support the launch, Cerebras will install Helios hardware within its own data centers and offer the combined service as a cloud option. This approach gives developers access to high-throughput processing without sacrificing system responsiveness, enabling businesses to run complex AI workflows without the capital expense of building their own infrastructure.