Blog October 6, 2025 Arm’s chiplet-based CSS platform is making custom AI silicon accessible to more companies, accelerating innovation beyond the world’s largest hyperscalers.

Share

As artificial intelligence (AI) workloads surge in complexity and scale, the traditional system-on-chip (SoC) model is encountering a trifecta of challenges: power inefficiency, performance bottlenecks, and lengthening time‑to‑market. Monolithic chip design – once the domain of hyperscalers – can no longer keep pace with the demands of modern AI infrastructure.

At the 2025 OCP Global Summit, Arm is spotlighting a shift: chiplet‑based innovation, powered by Arm Compute Subsystems (CSS) and Chiplet System Architecture (CSA), now opening doors for Tier‑1 silicon providers to build AI-optimized designs without needing the scale of a hyperscaler.

From Monoliths to Modular AI Silicon

For years, full custom SoCs dominated the high end of AI infrastructure; everything from compute blocks to memory controllers, interconnects, and accelerators was integrated on a single monolithic die. While this design offered tight control and performance advantages, it also carried steep trade‑offs:

• Rising power and thermal costs as process nodes push limits;
• Complexity in validation and verification of large, heterogeneous blocks; and
• Long lead times for design, tooling, and manufacturing. Enter chiplet‑based compute. By decomposing a system into smaller, specialized dies – compute, memory, I/O, accelerators, SoC architects and designers gain the ability to mix and match components, scale only what’s needed, and iterate faster. Until now, that modularity came with its own barriers: design fragmentation, lack of standardized interconnects, IP reuse challenges, and substantial upfront risk and cost.******

The Unlock: Arm’s CSS and CSA Model

Arm is closing those gaps via two foundational frameworks:

• CSS: These are pre‑validated, high‑performance IP building blocks – compute cores, AI accelerators, memory subsystems – with design, verification, and performance profiles already proven in real or emulated silicon. Using CSS means designers don’t have to invent every block from scratch or re‑validate what works; instead, they leverage established, optimized pieces.
• CSA: An open, standards‑driven architecture for how chiplets interconnect, communicate, and integrate across vendors. CSA defines electrical, physical, and protocol‑level compatibility so that IP from different sources – for example, accelerators from partner A, and memory dies from foundry B – can interoperate reliably on a shared platform. Together, CSS and CSA enable Tier‑1 silicon providers – companies such as Socionext, MediaTek, and others – to build custom, AI‑optimized chips that deliver performance comparable to hyperscaler designs, yet with lower risk, faster cycle times, and more flexibility. These providers can pick and choose compute blocks, accelerators, memory types, and integration paths according to specific workload needs, like vision models, inference engines, and multi‑tenant instances, rather than being locked into monolithic design trade‑offs.

The Role of OCP in Accelerating the Movement

The Open Compute Project (OCP) has long been a locus for open hardware collaboration, modularity, and efficiency – principles that align closely with the chiplet revolution. At the 2025 OCP Global Summit, Arm is demonstrating not just theoretical architectures but working examples of how CSS and CSA combinations are being used by Tier‑2 cloud service providers (CSPs), OEMs, and silicon vendors to future‑proof their AI infrastructure.

Key benefits that OCP partners are seeing include:

• The** flexibility** to customize silicon for region‑specific power, thermal, or reliability constraints.
• Lower total cost of ownership (TCO) through supply chain optionality, which is the ability to source chiplets or dies from multiple foundries, mix and match dies as volumes scale, rather than being tied to a single monolithic vendor.
• Faster time‑to‑market, since validated CSS blocks and standardized interconnect allow much of the design work to be “already done,” enabling more rapid prototyping, testing, and deployment.

Business Impact and What Comes Next

For AI infrastructure builders – whether Tier‑2 CSPs, OEMs, or silicon firms just stepping into AI work – the CSS and CSA approach offers real outcomes:

• Performance-per-watt improvements: Compute and memory are placed where they can be most efficient, without waste.
• Reduced design risk: By re‑using proven IP and relying on standard interconnects.
• Supply chain resilience: Modularity makes it more feasible to switch sources, scale die production, or choose preferred foundry nodes.”
• Speed: Design cycles shrink, allowing for faster iteration on AI models, feature sets, and deployment. This is more than silicon architecture; it’s a business lever for agility in the age of AI.

Learn More

At the 2025 OCP Global Summit, Arm will host sessions and technical briefings showcasing CSS and CSA in action. Whether you’re a silicon designer, infrastructure architect, or cloud provider exploring the future of AI hardware, this is your opportunity to see what’s possible beyond today’s constraints.

Arm’s mission is clear: to make custom AI silicon accessible to all, not just the hyperscalers, because the future of the datacenter depends on innovation at every scale.

Share

Article Text

Copy Text **

Any re-use permitted for informational and non-commercial or personal use only.

Editorial Contact

Arm Editorial Team

editorial@arm.com