Arm’s Agent CPU Evaluation Expands Beyond Token Throughput
Arm has proposed an evaluation methodology that measures CPU performance in agentic AI workloads, including tool calls, task orchestration, reasoning support, sandboxing and end-to-end execution. Arm has explored the approach with Fujitsu, Meta and OpenAI, but Ravi Malhotra said it is a methodology rather than an industry standard and should not use uniform weights across agent scenarios. Signal65 research commissioned by Arm modeled Arm AGI CPU as delivering about 1.99 times the average rack-level performance of AMD EPYC 9755 under the tested power budget, although the results were modeled rather than measured in a complete production rack.
Agent workloads make CPU orchestration and execution part of task performance, so infrastructure comparisons may need to cover end-to-end completion under latency and concurrency constraints rather than token generation alone.
过去两年，AI 基础设施的竞争几乎围绕 GPU 展开。模型能生成多少 Token、每秒能输出多少 Token、每瓦功耗能够支撑多少推理吞吐量，成为衡量算力效率的重要指标。
但随着 AI Agent 从聊天问答走向软件开发、企业自动化和复杂任务执行，一个新的问题开始浮现：如果模型生成了代码，却要花很长时间才能完成编译和测试；如果推理速度足够快，工具调用、数据库访问和任务编排却跟不上，那么更快的 Token 生成速度，究竟能带来多少实际生产力？
这也让 CPU 在 AI 基础设施中的作用开始被重新审视。
Arm 云 AI 解决方案架构总监 Ravi Malhotra 在一场技术分享中提出，智能体 AI 正在改变 CPU 性能的评价逻辑。过去，AI 性能主要围绕 Token 生成展开，而在 Agent 场景下，基础设施不仅需要支持模型思考，还必须让模型生成的内容转化为真实行动。
#Arm#Fujitsu#Meta#OpenAI#Signal65#AMD#Intel#Docker#Firecracker#gVisor#Arm AGI CPU#AMD EPYC 9755