AMD Says Agentic AI Brings New Attention To CPUs In The AI Data Centre

The rise of agentic AI is shifting the spotlight back onto CPUs in the data centre, with AMD arguing that processors are no longer just a supporting act but a critical part of orchestrating complex AI workloads.

In a blog post published by the chipmaker, AMD outlined how CPUs are taking on a greater role as AI deployments move beyond simple model training into multi-step inference workflows driven by autonomous AI agents.

AMD CEO and Chair, Dr Lisa Su, described agentic AI as “a new class of user” at the company’s Advancing AI event last June – referring to systems that are always active, continuously accessing data, applications and services to make decisions and complete tasks.

CPUs Step Into The Spotlight

While GPUs remain essential for the high-throughput parallel processing used in AI training and inference, AMD contends that modern AI deployments depend on balanced systems where CPUs, GPUs, networking and software each play distinct roles.

Within these environments, CPUs orchestrate workloads, manage memory and data movement and support the enterprise applications that run alongside AI models in production.

AMD likened the relationship between CPUs and GPUs to that of a head coach and a team of athletes. The CPU calls the plays, reacts to changing conditions and keeps all the players moving in the right direction, while GPUs specialise in performing one efficient part of the task at a time.

Training Versus Inference

The demands on CPUs differ depending on whether a system is training or running inference.

During training, GPUs do the heavy lifting, crunching large grids of data repeatedly so the system can learn. CPUs manage and feed data to GPUs, run the operating system, manage memory and schedule tasks.

As the bulk of AI work transitions to inference, AMD argued the CPU’s role becomes more intense. The CPU collects data, routes information, interprets results and decides final actions, with control, coordination and decision-making happening simultaneously.

“Architecture matters,” AMD wrote in the blog post, pointing to its chiplet design approach as giving the company flexibility to tune compute, I/O, memory bandwidth and power envelopes for different workloads.

Agentic AI Raises The Stakes

Agentic AI – artificial intelligence that can plan, decide and take actions with minimal human intervention – places new demands on CPUs, according to AMD.

In agentic workflows, CPUs spend more time considering results rather than simply returning a response. They often send problems back to GPUs for another pass with adjusted parameters before delivering a final result.

On top of existing duties, CPUs in agentic AI systems need to manage tool calls, API requests and memory queries while keeping GPUs busy. The rise of agentic AI drives increased CPU cycles as processors move data between agents, enterprise applications and data stores.

AMD argued the decisions CPUs make determine how GPUs are utilised, overall throughput and total cost of ownership for AI providers.

AMD EPYC Performance Claims

AMD pointed to recently published benchmarks to back up its case for CPU performance.

The company claimed a 5th Gen AMD EPYC CPU-based system is estimated to deliver up to 2.1x higher performance per core compared to Nvidia Grace Superchip-based systems. The same EPYC-based system is estimated to deliver up to 2.26x uplift on SPECpower, measuring operations per watt, against the same Nvidia system.

AMD also emphasised the x86 architecture advantage, noting that the vast majority of enterprise workloads already run natively across on-premises and cloud environments on x86 – without the refactoring, recompiling and multiple code bases that can come with adding Arm-based systems.

Building A Full-Stack AI Platform

AMD positioned its EPYC server CPUs as part of a broader system-level approach, working alongside AMD Instinct GPUs, AMD Pensando networking technologies and the AMD ROCm software stack.

The company argued that AI performance is increasingly defined at the system level, and that AMD is positioned to optimise everything from CPUs and GPUs to networking and software to maximise cluster-level performance per system watt.

Looking ahead, AMD flagged its next-generation EPYC CPUs, codenamed Venice, which are set to power the upcoming Helios rack-scale AI architecture. Venice is expected to extend performance, density and energy efficiency for AI and general-purpose workloads.

AMD concluded that AI is driving compute demand across the board and fuelling a global server refresh cycle, with EPYC processors providing the CPU foundation for customers looking to scale their AI infrastructure.

Last Updated on March 18, 2026 by Nick Ross