Arm recently unveiled the AGI CPU, a processor specifically designed to run AI agents, but Intel's leadership questions whether this specialized approach is actually necessary. The debate highlights a fundamental disagreement about how to optimize computing hardware for the next generation of artificial intelligence workloads. While Arm argues that existing x86 processors waste power and die space on features agents don't need, Intel contends that many of those features remain valuable and that current chips already perform well for agentic tasks .

What Makes Arm's New AI Chip Different?

Arm's AGI CPU is a 136-core processor designed from the ground up for running AI agents, which are software systems that can perform tasks autonomously using tools and APIs. The chip operates at 300 watts and strips away several features found in traditional server processors. Arm argues that conventional x86 processors, made by Intel and AMD, include unnecessary components that consume power without benefiting agent workloads .

Arm's design choices reflect this philosophy. The company removed simultaneous multithreading (SMT), a feature that allows a single processor core to handle multiple tasks at once. Arm also minimized Single Instruction, Multiple Data (SIMD) vector processing capabilities, which are used for certain types of mathematical operations. Instead of the 512-bit wide vector units found on Intel and AMD server chips, Arm's AGI CPU includes only 128-bit wide vector units. The chip delivers 6 gigabytes per second of memory bandwidth per core, more than double what Intel's comparable Clearwater Forest processors offer .

"We're focused on exactly and only what the agentic datacenter needs: performance, scale, and efficiency," said Mohamed Awad, executive vice president of Cloud AI at Arm.

Mohamed Awad, Executive Vice President of Cloud AI at Arm

Why Does Intel Disagree With This Approach?

Kevork Kechichian, who leads Intel's Data Center Group and previously served as executive vice president of Arm's Solutions Engineering team, isn't convinced that Arm's specialized design is necessary. He acknowledges that some of Arm's optimization points make sense, particularly the decision to minimize SIMD capabilities. For agent workloads focused on orchestration and data movement rather than complex mathematical operations, lighter SIMD engines could indeed be beneficial .

However, Kechichian disputes Arm's reasoning on other design choices. He points out that Nvidia, which also unveiled an agentic CPU platform called Vera, actually included SMT-like functionality in its design. Nvidia's approach uses "spatial multithreading," which splits each core's resources rather than time-slicing like traditional SMT. This suggests that if Arm had the technical option to include SMT, the company would have done so .

"My view is that, if they had the option, they would have put it in. They don't have the option, and none of the cores have SMT at Arm," Kechichian stated.

Kevork Kechichian, Data Center Group Chief at Intel

What Existing Intel Chips Already Do This?

Interestingly, Intel already offers processors that share many characteristics with Arm's AGI CPU. The company's Clearwater Forest processors pack 288 stripped-down cores with minimal SIMD extensions and 12 channels of fast DDR5 memory. These chips lack SMT and deliver the density and high core count that Arm emphasizes as critical for agent workloads. Kechichian notes that Clearwater Forest shares many qualities with Arm's new design, though Arm argues its product delivers superior memory bandwidth per core .

Intel also offers Sierra Forest and other efficiency-focused processors alongside its higher-performance Granite Rapids P-core Xeons. The company plans to offer Xeon 6+ configurations ranging from 288 cores at the high end down to the low 100s at the low end. However, Kechichian reveals that Intel doesn't see much demand for Xeon 6+ in agentic use cases. Instead, the chip is most popular in networking applications like packet processing .

How to Evaluate CPU Options for AI Workloads

• Memory Bandwidth Per Core: Compare how much data each processor core can access per second; Arm's AGI CPU delivers 6 gigabytes per second per core, while Intel's Clearwater Forest offers lower bandwidth per core due to its different core-to-memory ratio.
• Feature Necessity: Assess whether your specific workload actually benefits from features like SMT, boost modes, and SIMD extensions; agent frameworks may not need all the capabilities that traditional server workloads require.
• Power Efficiency: Evaluate sustained power consumption under typical workloads; Arm argues that x86 boost modes aren't sustainable across long periods, while Intel contends that its accelerators like QuickAssist remain relevant for compression and cryptographic tasks.
• Real-World Performance: Request benchmarks specific to your use case rather than relying on theoretical specifications; the jury remains out on whether Arm's optimization points actually translate to better agentic performance in practice.

Is This Really About AI Agents, or Something Else?

The debate between Arm and Intel reveals deeper questions about processor design philosophy. Arm's argument centers on the idea that purpose-built hardware optimized for specific workloads outperforms general-purpose processors. This philosophy has driven much of the AI acceleration market, where specialized chips like GPUs and tensor processing units (TPUs) have dominated AI training and inference .

However, Kechichian's skepticism suggests that the industry may be overestimating how different agentic workloads actually are from existing server workloads. Many of the features Arm removed, like SMT and broad SIMD support, have proven valuable across diverse computing tasks. The fact that Nvidia included SMT-like functionality in its Vera CPUs suggests that even companies building agentic platforms see value in these capabilities .

Kechichian isn't ruling out the possibility that demand for agentic workloads will eventually justify specialized CPUs. He simply argues that the case hasn't been made yet, and that existing processors already handle these tasks reasonably well. As enterprises and hyperscalers begin deploying AI agents at scale, real-world performance data will ultimately settle this debate .

The outcome matters because it will influence how technology companies invest in processor design over the next several years. If Arm is right, we'll see a wave of specialized agentic CPUs from multiple vendors. If Intel is right, general-purpose processors with optional efficiency cores will continue to dominate, and companies like Arm may struggle to justify the investment in purpose-built designs.