AMD highlights how CPUs power end-to-end agentic AI workflows beyond traditional GPU focus

AMD Redefines AI Infrastructure Around Workflow Stages

AMD is challenging the conventional wisdom that AI infrastructure begins and ends with GPUs. Instead, the chipmaker argues that agentic AI workflows demand a portfolio approach to CPUs, with different processors optimized for distinct stages of how AI agents operate 1

1

2

2

. When an agent receives a task, it doesn't simply generate a response. It interprets intent, retrieves context, plans steps, calls tools, applies policy, runs sandbox code, executes transactions, observes outcomes, and returns results. Each step represents a different workload with unique computational needs, making end-to-end agentic AI workflows far more complex than traditional AI deployments.

Diverse Compute Requirements Across the Agentic Pipeline

The AMD EPYC CPU portfolio addresses three critical workflow stages with specialized processors. For orchestration and tool execution, where multiple agents simultaneously run sandbox code or query databases, core counts matter more than clock speed. The 5th Gen AMD EPYC server CPUs deliver up to 192 cores and 384 threads, while the upcoming "Venice" processors will push that to 256 cores and 512 threads 2

2

. For enterprise application integration, the AMD EPYC 9005 family provides 8 to 192 cores with up to 640GB/s of memory bandwidth, scaling to handle varied incoming requests. When agents need reasoning capabilities through LLMs running on GPUs, host-node CPUs require high single-core performance to keep accelerators fully utilized. The AMD EPYC 9575F processor delivers 64 cores capable of running at up to 5GHz, ensuring GPU clusters generate maximum tokens 1

1

.

Enterprise Adoption Challenges and Legacy Infrastructure

AMD identifies two patterns creating enterprise adoption challenges for agentic AI. Many organizations standardize CPU infrastructure purchases around legacy CPU specifications such as 16- and 32-core processors. However, agentic workflows need higher core counts for some stages and higher frequencies for others, requiring flexibility rather than rigid standardization 2

2

. The mindset must shift from a single CPU standard to a portfolio matched to workflow requirements. Once employees gain the ability to build and deploy their own agents, agentic adoption accelerates rapidly, creating a multiplier effect on IT infrastructure that extends to databases, enterprise resource planning platforms, customer relationship management systems, business intelligence tools, identity management, and inference servers 1

1

.

AI Infrastructure Planning for CIOs

IT leaders treating agentic AI as a monolithic problem with one GPU strategy or a one-size-fits-all CPU approach will likely encounter scaling challenges. AMD's position centers on matching infrastructure to how agentic AI actually works across its many stages and workloads. The company's strategy encompasses EPYC CPUs for high-frequency and high-density compute, AMD Instinct accelerators for AI inference and training, and Pensando networking to move data predictably 2

2

. As agents become greater users of existing IT infrastructure, planning teams need to anticipate what happens when usage dramatically increases. The question worth asking isn't how many CPUs or GPUs a business needs for agentic AI, but whether infrastructure matches the way agentic AI operates with different compute needs at each stage. Organizations that map these stages early and choose the right compute profile for each workflow component will be better positioned for speed and efficiency as they scale their agentic deployments.