SambaNova and Intel have unveiled a new heterogeneous AI inference architecture combining Intel Xeon 6 processors with SambaNova RDUs and GPUs to accelerate agentic AI workloads. The solution targets coding agents and other autonomous applications requiring rapid tool execution and low-latency inference at scale. The design will be available to enterprises and cloud providers in H2 2026.

This architecture addresses the limitations of GPU-only stacks by assigning specific roles to each component: GPUs handle the parallel prefill phase, SambaNova RDUs manage high-throughput decode operations, and Xeon 6 processors serve as both host and action CPUs for tool execution and code compilation. SambaNova claims Xeon 6 delivers over 50% faster LLVM compilation compared to Arm-based server CPUs and up to 70% faster vector database performance than competing x86 processors.

The collaboration marks a significant shift toward heterogeneous computing for production AI workloads. By leveraging Xeon 6's mature x86 ecosystem alongside purpose-built inference accelerators, the solution promises cost-efficient scaling for organizations deploying thousands of simultaneous coding agents. This approach enables enterprises to run complex agentic workflows within existing air-cooled data centers while maintaining compatibility with current software infrastructure.