AMD and Cerebras Unveil Helios AI Rack-Scale Platform Combining EPYC CPUs and Wafer-Scale Engines
AMD and Cerebras are partnering to build a disaggregated inference platform, pairing AMD’s Helios rack-scale systems with Cerebras’ Wafer-Scale Engines to split the prefill and decode stages of LLM inference.
AMD and Cerebras Systems have announced a collaborative effort to develop a disaggregated inference platform that marries AMD’s Helios rack-scale infrastructure with Cerebras’ Wafer-Scale Engine (WSE) technology. According to Tom's Hardware, the system is designed to partition inference workloads by stage: AMD’s Instinct MI400-series accelerators, housed within Helios racks alongside EPYC processors, will handle prompt processing and large context windows. Cerebras’ WSE, which utilizes on-chip SRAM rather than HBM4, will take over the memory-bandwidth-bound token generation stage. This division of labor aims to optimize performance for latency-sensitive agentic workloads.
The technical rationale centers on the disparate memory requirements of modern large language models. While GPUs are effective for the compute-intensive prefill phase, the decode phase is heavily constrained by memory bandwidth. By leveraging the high SRAM bandwidth inherent in the Cerebras WSE, the companies claim they can achieve significantly higher interactivity. According to The Register, Cerebras CEO Andrew Feldman noted that the WSE’s SRAM-based architecture allows for output speeds that can exceed 2,000 tokens per second, providing a distinct performance profile compared to traditional HBM-based GPU clusters.
This disaggregated approach mirrors the design principles seen in other recent rack-scale architectures, such as the Nvidia-Groq integration. However, the AMD-Cerebras collaboration inverts the specialization. Whereas other designs might utilize specific hardware for the context/prefill stage, this platform assigns the compute-heavy prefill to AMD’s Instinct GPUs and the memory-intensive decode to the WSE. This specialization is projected to deliver up to 5x higher tokens per second per watt (T/s/W), according to both companies, though they have yet to disclose granular performance benchmarks or the specific interconnect methodology linking the two compute platforms.
The scale of the hardware required for large models is a primary differentiator for this partnership. According to The Register, while serving a trillion-parameter model like Kimi K2.5 might require thousands of Groq LPUs, the combined AMD and Cerebras architecture is expected to handle similar workloads with only a few dozen WSEs. This efficiency stems from the unique wafer-scale integration of the Cerebras chips, which avoids the bottlenecks typically associated with scaling across traditional GPU interconnects for memory-bound tasks.
Cerebras plans to integrate these AMD Helios systems into its existing data center footprint. The combined offering is slated to become available through Cerebras Cloud in the second half of 2026. During a press conference following the Advancing AI keynote, AMD CEO Lisa Su indicated that this partnership aligns with AMD’s broader strategy of fostering an open ecosystem for workload-specific acceleration. Su suggested that further disaggregated compute solutions could be expected in the future, as the company explores various technologies that complement the Helios rack architecture.