AMD presented the direction of changes in data center infrastructure driven by the spread of agentic AI. It emphasized the importance of rack-level CPU performance and processing capabilities, moving away from a simple GPU-centric structure.

AMD announced on the 10th that the role of CPU-based infrastructure in agentic AI environments is expanding and rack-scale performance is emerging as a key metric.

Agentic AI consists of various service layers as well as model inference. Multiple systems, such as orchestration, databases, web services, APIs, caches, and middleware, operate simultaneously.

Most of these systems are highly dependent on CPU resources, and their infrastructure scales according to the number of concurrently running agents.

AMD explained that in such environments, CPU processing power, rather than GPU performance, is the factor determining overall system scalability.

AMD proposed 'rack-unit throughput' as a standard for evaluating data center performance. This is because actual data centers are built under constraints such as power, cooling, and space.

Accordingly, they stated that “how much work can be processed in a 100kW rack” is more important than single-chip performance.

According to AMD, under the same conditions, the EPYC 9965 was analyzed to have a rack throughput of approximately 2.37 times that of NVIDIA Vera-based systems and approximately 1.6 times that of Intel Xeon.


It was explained that the next-generation EPYC 'Venice' is expected to further widen that gap.

AMD emphasized that EPYC-based systems provide high core density per rack.

Currently, EPYC 'Turin'-based systems can support more than 27,000 CPU cores per rack, and next-generation products are expected to expand to more than 36,000.

AMD explained that this high-density structure provides the foundation for handling more services and agents within the same power budget.

AMD also emphasized that the infrastructure can be built in an existing x86-based environment without a separate new architecture.

This was cited as a factor that reduces the burden of infrastructure transition while maintaining software compatibility.

AMD stated that in agentic AI environments, overall system balance including CPU, GPU, and network, as well as rack-level efficiency, are more important than the performance of a single component.