The rise of agentic AI is reshaping how modern data centers are designed, with CPUs now taking on a more critical role alongside GPUs in AI infrastructure.
At its Advancing AI event, AMD CEO Lisa Su described agentic AI as a new class of systems capable of continuously interacting with data, applications, and services to make decisions and execute complex tasks. Unlike traditional AI workloads, these systems operate as persistent agents, increasing the complexity of inference workflows.
This shift is driving renewed focus on CPUs, which are responsible for orchestrating these increasingly multi-step AI processes.
CPUs Move From Support to Coordination Layer
In traditional AI pipelines, GPUs handle parallel processing tasks such as training neural networks and running inference workloads. However, as AI systems become more dynamic, CPUs are increasingly responsible for coordinating the entire pipeline.
Within modern AI clusters, CPUs handle scheduling, memory management, data preparation, and I/O operations. They also manage control flow and ensure that GPUs remain fully utilized.
The relationship can be simplified as orchestration versus execution. GPUs process data at scale, while CPUs ensure that data arrives correctly, workflows remain synchronized, and system resources are efficiently allocated.
Inference Workloads Are Changing the Balance
The growing emphasis on inference is one of the main reasons CPUs are becoming more important.
During training, workloads are largely predictable and GPU-heavy. CPUs act primarily as support, feeding data and managing system operations. In contrast, inference in agentic AI environments involves multiple steps, decision-making loops, and interactions with external systems.
This shifts more responsibility to the CPU, which now handles tasks such as routing outputs, interpreting results, managing API calls, and coordinating iterative processing between models.
As a result, CPU utilization increases significantly in production AI environments, particularly those deploying autonomous or semi-autonomous agents.
Performance and Efficiency Remain Key Differentiators
AMD positions its EPYC processors as a foundation for these evolving workloads, emphasizing both performance and energy efficiency.
According to AMD, a 5th-generation EPYC CPU-based system can deliver up to 2.1 times higher performance per core compared to systems based on Nvidia’s Grace Superchip. The same configuration is also estimated to achieve up to 2.26 times better performance per watt, based on SPECpower benchmarks.
The company also highlights the advantage of x86 architecture, which benefits from a mature software ecosystem and broad compatibility across enterprise workloads without requiring code refactoring.
Balanced Systems Become the Priority
The broader takeaway is a shift toward balanced AI infrastructure.
Rather than focusing solely on GPU performance, data center operators are increasingly optimizing entire systems, including CPUs, networking, and software stacks. CPUs play a central role in this approach by enabling efficient GPU orchestration and supporting enterprise applications running alongside AI workloads.
AMD’s strategy reflects this direction, combining EPYC CPUs with Instinct GPUs, Pensando networking, and the ROCm software platform to deliver integrated AI systems.
What Comes Next
Looking ahead, AMD is preparing its next-generation EPYC processors, codenamed “Venice,” which are expected to power future AI platforms such as the “Helios” rack-scale architecture.
As AI workloads continue to evolve, the role of CPUs is expected to expand further, particularly in managing complex, real-time, and multi-agent systems.
The shift underscores a broader industry trend: AI performance is no longer defined by accelerators alone, but by how effectively the entire system operates as a cohesive unit.