AMD Launches 256-Core Venice CPU for Agentic AI Era

AMD has launched the EPYC Venice server CPU featuring up to 256 cores. CEO Lisa Su positions it as the primary processor for Agentic sandboxes and AI hosts. AMD argues that as AI transitions from chatbots to Agentic AI, the compute focus in data centers will shift back from GPUs to CPUs. Future data centers will segment into distinct CPU server categories tailored for general-purpose and AI-specific workloads. Utilizing a chiplet strategy to deliver a broad range of SKUs, AMD demonstrated that Venice offers superior energy efficiency and performance compared to competitors like the NVIDIA Vera and ARM AGI.

已确认

- The Venice server CPU family scales up to 256 cores, leveraging a chiplet architecture with 8 compute dies (m1, m2).

- AMD CEO Lisa Su emphasized that Venice is designed for Agentic sandboxes and AI host scenarios (m1).

- AMD predicts that the Agentic AI era (post-2025) will drive new demand for CPU servers, leading data centers to split into three workload types: general-purpose CPU servers, Agentic CPU servers, and others (m4, m6).

- AMD noted that while past chatbot workloads typically required a 1:4 CPU-to-GPU ratio, Agentic AI will alter this dynamic, making CPUs a primary growth driver once again (m6).

- In live benchmarks, AMD claimed Venice achieves 1.2x the single-core performance and 2.2x the throughput of NVIDIA Vera (m7).

- Venice demonstrated up to a 3.3x improvement in energy efficiency during server demos, and up to 2.8x higher agent performance per watt in agentic CPU servers (m5).

- OpenAI's Head of Infrastructure also appeared on stage to discuss collaborations related to Helion (m2).

为什么重要

AMD is attempting to redefine the compute landscape of the AI era. With the market heavily focused on GPUs over the past two years, AMD's core thesis is that Agents will require significantly more CPU power for orchestration and sandbox execution. This shift will unlock new growth avenues for the EPYC product line. By directly benchmarking against the NVIDIA Vera platform, AMD is also signaling that it will no longer avoid a full-stack AI infrastructure rivalry with NVIDIA.

2026-07-24 ~ 2026-07-24 · 7 related posts

Full story(4 episodes)→

Primary sources