INTERVIEW In recent weeks, the likes of Nvidia and Arm have revealed CPUs designed expressly to run AI agents like OpenClaw.

Kevork Kechichian, who runs Intel's Data Center Group and served as executive vice president of Arm's Solutions Engineering team until last summer, isn't so sure this "new" kind of CPU is really what hyperscalers or enterprises actually need.

His comments came just days after Arm unveiled its full processor design, a chip called the AGI CPU, which it proposes as an agentic AI processor. Nvidia showed off its own agentic compute platform, powered by its in-house Vera CPUs, a week earlier.

After years of GPUs and AI accelerators dominating headlines, CPUs are back in the limelight because those agentic frameworks, tools, API calls, and AI-generated code snippets need to run on something, and it's not GPUs.

Speaking at the Arm Everywhere event in San Francisco last week, Arm EVP of Cloud AI Mohamed Awad made the case that existing x86 processors weren't designed to run agents, and that their boost modes, simultaneous multithreading (SMT), specialized accelerators and other legacy features that work for today's workloads only served to consume die area and drive up power consumption.

"When you increase the frequency, what else do you increase? Power. That's a problem. These boost modes are not sustainable across long periods of time. They're not sustainable across a chip," Awad claimed.

Naturally, Arm argues its 300-watt, 136-core chip avoids those problems.

"We don't support Lotus Notes, we just don't do it," Awad said in an apparent reference to x86 real mode. "We're focused on exactly and only what the agentic datacenter needs, performance, scale, and efficiency."

The cores Arm uses in the AGI are also surprisingly light on Single Instruction, Multiple Data (SIMD) features compared to the AVX extensions found on modern x86 server processors. Arm's chip features a pair of 128-bit wide vector units, compared to the 512-bit wide vectors supported on most Intel and AMD server chips.

Awad went out of his way to pitch the chip's lack of SMT, which you might know as hyperthreading, as a benefit rather than a negative.

"What happens when you do multithreading? You throw two jobs at the same core, that's how they get to a high thread count," he said. "The reality is that your I/O and your bandwidth don't double, so you've just moved the bottleneck elsewhere."

SMT is great unless you don't have it, in which case it's terrible

Whether or not the optimization points highlighted in Arm's AGI CPU announcement are the ones that actually matter for agentic performance, the jury is still out for Intel's Kechichian.

One area where he can see the logic is on SIMD.

"If you look at the workloads, it's just mostly traditional data movement types of things; orchestration," he said. "That's one area where not having heavy SIMD engines is a good thing."

He also acknowledges that there are features in current CPUs, both Arm and x86, that you don't necessarily need for agent frameworks. However, he argues that many of the accelerators Intel has developed over the past several years remain relevant – for example, QuickAssist, which is designed to speed up compression, decompression, and cryptographic workloads.

Kechichian is also less than convinced by Arm's case against SMT. "While Renee talked about non-SMT and optimization, a week before, Jensen showed another CPU which has SMT."

Nvidia's Vera CPUs feature 88 of its custom Arm-based Olympus cores, which include what the GPU giant calls "spatial multithreading." As Nvidia explains it, the tech essentially splits each core's resources down the middle rather than doing time slicing like any other x86 chip with SMT would.

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

That said, it's also important to understand that some workloads have always benefited from SMT more than others. There is a reason that IBM is still shipping new Power CPUs with four or even eight threads per core.

But because of this, Intel and AMD have long made it easy to turn SMT on or off in BIOS settings, at least for the parts that support multiple threads per core in the first place.

By Arm's logic, Intel already has an agentic CPU

Alongside its Granite Rapids P-core Xeons, Intel also has its Sierra Forest and Clearwater Forest processors that pack in plenty of its ultra-efficient cores.

Clearwater Forest in particular shares many qualities with Arm's AGI CPU. It's got 288 stripped-down cores with minimal SIMD extensions and 12 channels of fast DDR5 memory.

"It has the density, it has the high core count, and it also lacks SMT," Kechichian said.

Asked about the similarities between Arm's product and Intel's Clearwater Forest, Awad argued that the parts were really designed around maximizing compute density, citing the memory bandwidth per core and calling into question relative performance of Intel's efficiency cores.

While it's true that Arm's 136-core parts deliver 6 GB/s of memory bandwidth per core, this is largely down to the ratio of compute to memory. In fact, it is common to see lower core count parts with large caches favored for memory-bound workloads like computational fluid dynamics.

Fewer cores hanging off the same memory subsystem usually, but not always, translates to higher bandwidth per core.

Compared to Intel's top-specced Clearwater Forest parts, Arm's CPU offers more than twice the bandwidth per core.

We don't have the full Xeon 6+ SKU list just yet, but Kechichian tells us the part will be offered in configurations ranging from 288 cores at the high end to the low 100s at the low end. In a 136-core vs 136-core scenario, Arm's lead would likely be significantly smaller.

Despite checking many of the same boxes as Arm's AGI CPUs, Kechichian tells us Chipzilla does not see much demand for Xeon 6+ in agentic use cases. Instead, we're told the chip is most popular in networking applications like packet processing.

Kechichian isn't ruling out the possibility that demand for agentic workloads will come in time. ®