AMD: Future is about finishing workflows,not winning token races

AMD said its Ryzen AI Halo platform is designed for the emerging era of AI agents, arguing that overall workflow completion — not raw AI token generation — has become the key benchmark for enterprise artificial intelligence.

The chipmaker said AI workloads are increasingly shifting from simple chatbot responses to autonomous agents capable of researching, retrieving information, generating content and validating results.

Unlike traditional AI benchmarks that focus on tokens per second, AMD said agentic AI should instead be measured by end-to-end workflow completion time and cost per completed task.

To demonstrate the difference, AMD developed the Hermes Executive Presentation Agent (HEPA), a local AI workflow that generates executive presentations, reports, charts and supporting documents using a local knowledge base.

According to AMD, systems powered by the Ryzen AI Max+ 395 processor completed the workflow 15 percent faster than NVIDIA’s DGX Spark Grace Blackwell platform while delivering 34 percent faster CPU orchestration and 27 percent lower cost per completed workflow.

AMD attributed the gains to its balanced architecture, which combines CPU, GPU and unified memory to handle different stages of an AI workflow. While GPUs generate AI responses, the company said CPUs perform much of the orchestration work, including document parsing, optical character recognition (OCR), embedding, retrieval and validation.

The company said seven of the eight stages in its test workflow relied primarily on CPU performance, making overall system balance more important than GPU speed alone.