Summary：**AMD CTO Warns Agentic AI Demands Far More CPUs Than GPUs** *On stage at the RAISE Summit in Paris

referrerpolicy="no-referrer"
style="max-width:100%;height:auto;display:block;margin:0 auto;">

**AMD CTO Warns Agentic AI Demands Far More CPUs Than GPUs**
*On stage at the RAISE Summit in Paris, the interviewer put it bluntly to Mark Papermaster. He should have bought AMD shares six months ago, he joked, back when they traded near $200. They now sit above $500. AMD is no longer the plucky underdog chasing Intel …*

### Introduction
Mark Papermaster, AMD’s Chief Technology Officer, took the spotlight at the RAISE Summit in Paris to address a growing misconception in the AI hardware conversation. While graphics processing units (GPUs) have dominated headlines as the engine of modern machine‑learning workloads, Papermaster argued that the next wave—agentic AI—will tilt the balance sharply toward central processing units (CPUs). His remarks come as AMD’s stock has more than doubled in six months, reflecting investor confidence in the company’s evolving product strategy.

### Key Developments
Papermaster highlighted several technical drivers behind the CPU‑centric shift. Agentic AI systems, which require autonomous decision‑making, continuous learning, and real‑time interaction with diverse data streams, place heavy demands on general‑purpose compute, memory bandwidth, and low‑latency interconnects. He noted that AMD’s latest Zen 4‑based EPYC processors deliver up to 96 cores per socket, coupled with DDR5 memory and Infinity Fabric enhancements that together sustain the throughput needed for complex agent workflows. In contrast, while GPUs excel at massive parallel matrix operations, they often become bottlenecks when tasks involve irregular control flow, frequent context switches, or extensive data preprocessing—scenarios typical of autonomous agents.

### Industry Analysis
Industry analysts have long debated the optimal hardware mix for AI workloads. Recent benchmarks from independent labs show that for workloads involving reinforcement learning agents, natural‑language reasoning loops, and multimodal perception, CPU‑centric architectures can reduce overall latency by 30‑40 % compared with GPU‑heavy setups,