Computing Architectures for AI Workloads Face a Dilemma

A fork in the road is emerging in computing architectures that could upend the pecking order of today’s leading chipmakers.

For nearly half a century, the dominant architecture for the CPU was the x86, created by Intel. Most of the world’s computers, servers and data centers use the x86 architecture. Since it has been around since 1978, introduced as part of the Intel 8086 processor, this architecture ensures that older software written for earlier x86 processors can run on newer x86 CPUs. Intel and AMD are the two largest users of x86.

But a rising star is British chip designer Arm, which uses a different CPU architecture not compatible with x86. It is simpler, customizable and energy efficient, enabling it to garner a 99% market share in mobile devices because it’s better for battery-powered gadgets. In contrast, x86 is meant for higher-performance computing.

When Amazon Web Services (AWS) introduced its Arm-based processor Graviton in 2018, the duopoly of Intel and AMD took notice, according to Steve McDowell, CEO of Nand Research, writing in Forbes. Graviton was the first custom chip deployed at scale by a major cloud provider.

“Amazon didn’t intend to change the microprocessor market,” he said. “The company first built Graviton to take control of its roadmap, a way for the notoriously bottom-line-focused company to reclaim the margin it felt it was sacrificing to Intel and AMD.” However, “what it did instead was start a movement.”

AWS’ Graviton was the “first serious alternative” to the dominant x86-based server processors in the industry, McDowell added. It also opened the door to other tech giants thinking the same thing: Microsoft, Google, Apple, Nvidia, Meta, Qualcomm, Samsung, Huawei and Alibaba followed suit with their custom Arm-based chips.

Last October, fierce chip rivals Intel and AMD joined forces to try to stem the tide towards Arm and other architectures, such as the open source RISC-V. The two chipmakers formed the x86 Ecosystem Advisory Group to formalize standards as the industry moves toward artificial intelligence (AI) workloads.

Joining them are the who’s who in tech: Microsoft, Google Cloud, Dell, Oracle, Broadcom, Lenovo, HP, HPE, Red Hat and luminaries Linus Torvalds (Linux creator) and Tim Sweeney (CEO of Epic Games and creator of Unreal Engine).

“X86 has been the de facto standard in the industry for the last four decades [with the] broadest compatibility,” said Justin Hotard, executive vice president and general manager of the data center and AI group at Intel, at a recent conference. Most software is developed and runs largely on x86, whether it’s at the edge, on PCs, laptops or in the data center.

It is also the CPU architecture of choice for AI frameworks because it works with older software. That means x86 “ensures your investments are future-proofed while being backward compatible,” Hotard said.

Both CPUs and GPUs are used for AI workloads. While CPUs can technically handle AI workloads, GPUs are the preferred choice because these can better handle the massive computational demands of training and running complex AI models. CPUs handle general-purpose computing tasks.

The advisory group’s goal is to “make it easier and simpler for the entire ecosystem in the AI era to continue to use and benefit from x86,” Hotard said.

Arm Chief Commercial Officer Will Abbey, who recently spoke with PYMNTS, said his company welcomes the competition. “We’re a big believer in choice.” However, he declined to comment on rumors that Arm is looking to manufacture chips this year.

Hotard said going forward, x86 chips are also going to be more energy efficient. “AI is creating an unquenchable demand for energy. Research shows data center energy consumption is going to double, at least, in the next four years. We can’t make the world a better place with AI if we don’t make the world a greener place powered by AI.”

Energy efficiency is a critical priority for our investments in the ecosystem at Intel, he added. These include the Intel Core Ultra chip and the Xeon family of processors.

But Suren Nihalani, senior software engineer at Zoox, the Amazon-owned robotaxi company, believes the x86 advisory group was formed too late. They have “already lost the battle to Arm because of Apple,” he said in an interview with PYMNTS. As such, the working group is merely “a distraction.”

In 2020, Apple switched the chips in its Mac computers from Intel to its own custom M1 chips, using Arm designs. These chips are now in the iPhone and other Apple devices.

Notably, Apple was a co-founder of Arm in 1990 (then called Advanced RISC Machines Ltd.), joined by VLSI Technology (now NXP Semiconductors) and Acorn Computers. In 1993, Apple launched its now-defunct Newton handheld device using Arm architecture.

The post Computing Architectures for AI Workloads Face a Dilemma appeared first on PYMNTS.com.