Arm Holdings plc is working to deliver high-efficiency infrastructure standards for artificial-intelligence data centers to help address rising energy demands in conjunction with the Open Compute Project.

To address this challenge, Arm today announced that it will support companies in developing the next phase of purpose-built silicon and packaging for converged infrastructure. This approach combines co-designed capabilities that integrate compute, memory and networking interfaces on the same system-on-a-chip, enabling AI data centers to deploy denser, more efficient compute.

Traditional data centers rely on separate server boards that connect central processing units, graphics processing units, memory and networking components across a rack. In contrast, SoC designs consolidate these elements into a single tightly integrated package, reducing power draw and latency while enabling higher compute density for AI workloads.

To power the next generation of converged datacenters, Arm is contributing its new Foundation Chiplet System Architecture specification to the OCP and broadening the Arm Total Design ecosystem.

The Open Compute Project is a global industry organization that designs and shares open-source hardware configurations for data-center technologies and infrastructure. Its goal is to promote collaboration between companies, engineers and users to share ideas, specifications and intellectual property to advance hardware design. OCP is known for encouraging modular hardware components, an approach that aligns with emerging chiplet-based design.

Chiplets are modular integrated circuits that break down larger, more monolithic processor designs into components using a “Lego-like” approach, featuring smaller, specialized chips. This approach to modularity allows for the creation of more complex SoC designs built on more power-efficient silicon optimized for a specific task, providing flexibility and scalability for bespoke configurations.

“The converged AI data center won’t be powered by individual commodity chips,” said Mohamed Awad, senior vice president and general manager of infrastructure business at Arm. “To increase density of these systems, advanced purpose-built silicon is required.”

The FCSA builds on Arm’s earlier Chiplet System Architecture work and aims to address industry demands for vendor-neutral requirements. It provides a common set of standards for chiplet system and interface designs, enabling interoperability at scale irrespective of CPU architecture.

“The idea is when you go build that next AI accelerator SoC, you could ideally be able to leverage chiplets from other partners,” Eddie Ramirez, Arm’s vice president of marketing, infrastructure business, told SiliconANGLE in an interview. “You don’t have to design every single chiplet yourself.”

Ramirez explained that the growing demand for AI compute resources has led to a dramatic increase in rack density within data centers. At the same time, the integration point for compute has shifted from the server board to the SoC package itself. Power consumption has become one of the biggest concerns.

“If you look at these AI racks, for example, they’re now 10 times more power hungry than an average compute rack was five years ago,” he said.

In the past, a data-center rack held 30 to 40 servers and drew about 12 to 15 kilowatts. Today’s AI racks can pack 72 to 100 servers and pull as much as 100 kilowatts. By implementing chiplet-based designs, AI accelerators can achieve better performance per watt, helping to reduce operating costs.

Arm expands its Total Design ecosystem

To bring advanced chiplet designs to the industry, Arm announced an expansion of the Arm Total Design ecosystem. Announced two years ago, ATD has grown from 13 partners at launch to 36 partners in the silicon and chipset design industry. Today, the company is announcing 10 more partners for the program.

New partners include Astera Labs Inc. for input/output chiplets for memory and PCIe standards, Rebellions Inc. for AI accelerators and leading application-specific integrated circuit leader Marvell Technology Inc. Other additions are AIchip, ASE, CoAsia, Credo, Eliyan and Insyde Software. Ramirez said the combined expertise in packaging, interconnect and system integration will help drive the next phase of standards and accelerate chiplet innovation across the silicon-design lifecycle.

“The hope is that we can bring the cost down and shorten the design time of these really complex SOCs,” Ramirez said.

To lead this reconfiguration of infrastructure by the industry, Arm announced that OCP has appointed Arm to its board of directors. The company will join AI software and silicon manufacturing giants, including Meta Platforms Inc., Google LLC, Intel Corp. and Microsoft Corp., to assist with aligning the future toward interoperable design for AI datacenters.

Image: Arm