Arm Chiplet System Architecture Makes New Strides in Accelerating the Evolution of Silicon

The first public specificiation for the Arm Chiplet System Architecture is now available, with over 60 companies contributing to its development.

By Eddie Ramirez, VP, Infrastructure Business, Arm

AI has the potential to steer a new industrial revolution, permeating all markets unlike anything we’ve seen before. To get there, we must be able to address the wide variation in AI workloads across a diverse set of markets. For example, the appetite for latency changes depending on application. The response times for a system designed to help researchers simulate proteins will be different to the response times required for a system built to operate in a passenger vehicle. This wide range of compute requirements means we need to provide more than one type of compute solution; each optimized for specific market needs. The increasing demand for custom silicon, combined with the costs and complexities of silicon production, is driving the trend toward wider adoption of chiplets.

Video: What are chiplets explainer

Advancing the chiplet ecosystem through standardization and collaboration

The reuse of specialized chiplets to create multiple custom systems-on-chip (SoCs) can deliver systems with better performance and lower power consumption, at lower overall design cost compared to monolithic chips. However, without industry-wide standards and frameworks, variations in chipsets could lead to compatibility issues that ultimately slow innovation. To address this fragmentation, last year Arm introduced the Chiplet System Architecture (CSA). The CSA provides a set of system partitioning and chiplet connectivity standards that have been co-developed with the ecosystem, aligning the industry on the foundational choices of building chiplets. Using the CSA, new chiplet designs can be created with confidence that they can be adapted and reused in any compliant system, expediting chiplet-based system innovation while reducing the risk of fragmentation.

CSA’s first public specification is now available, supported by 60 leading players

Today, it’s my pleasure to share a significant milestone for CSA: The first public specification is now available. Representatives from over 60 companies are now engaged with the CSA, contributing to and applying standards in silicon strategies across multiple market segments. These companies include ADTechnology, Alphawave Semi, AMI, Cadence, Jaguar Micro, Kalray, Rebellions, Siemens, Synopsys, and others.

The breadth of engagement from these innovative technology companies forms the foundation of an Arm-based chiplet ecosystem set to revolutionize system design, making SoCs more flexible, accessible, and cost-effective while reducing the risk of fragmentation. With a public spec now available, designers have a shared understanding of how to define and connect chiplets into composable SoCs that can address the variance in AI workloads and ensure silicon is fit for specific markets.

Several vendors engaging with the CSA are already building solutions as part of Arm Total Design, an ecosystem dedicated to frictionless delivery of custom silicon powered by Arm Neoverse Compute Subsystems (CSS). To date, Arm Total Design has seen success in the deployment of chiplet-based compute subsystems that enable market-specific strategies including:

• Tailoring AI workloads for diverse markets: Alphawave Semi has customers that require performant chips for AI workloads, including networking, edge computing, storage, and security. By combining a chiplet powered by Arm Neoverse CSS with proprietary I/O dies, Alphawave Semi can use AMBA CHI C2C to connect accelerators tailored to the specific needs of each market. Custom SKUs for specific markets are derived from a standard base, amortizing the cost of the compute die whilst maintaining the flexibility to build multiple systems.

• Revolutionizing large-scale AI training and inference workloads: ADTechnology, Samsung Foundry, Rebellions and Arm have combined technologies to create an AI CPU chiplet platform for training and inference of large-scale AI workloads in the datacenter, with an estimated 2-3x efficiency advantage for GenAI workloads (Llama3.1 405B parameter LLMs). This multi-vendor chiplet platform combines Rebellions’ REBEL AI accelerator with coherent NPU’s using AMBA CHI C2C interconnect, and is built with a Neoverse CSS V3-powered compute chiplet from ADTechnology which can now be implemented with Samsung Foundry 2nm Gate-All-Around (GAA) advanced process technology as a result of the standardization work done with the CSA to date.

The role of chiplets in custom silicon for growing AI workloads

These are just a couple examples of how CSA can help address the wide variation of workloads being driven by AI across all markets, from infrastructure to automotive and consumer technologies. We believe the Arm-based chiplet ecosystem is uniquely positioned to meet the challenge of growing AI demands in all markets, leveraging the flexibility of the Arm compute platform, the seamless communication enabled by standards like AMBA CHI C2C, and the integration enabled by CSA. As the ecosystem around CSA continues to grow, so does the collaboration on standards and the impact we can make as an industry to significantly reduce fragmentation and enable faster development and deployment of custom silicon solutions.

To access the first public CSA spec at no cost, please visit this link. If you’re attending Chiplet Summit in Santa Clara, CA this week, I’ll be covering this topic and more in a keynote on January 22 titled Meeting the Challenge of Diverse AI Workloads with Chiplets.