Arm releases first public specification for chip system architecture

Eddie Ramirez, Vice President of Infrastructure Business Group at Arm, stated, "Artificial Intelligence (AI) holds immense potential to spearhead a new industrial revolution, with its market penetration reaching unprecedented depths and breadths. To achieve this ambitious goal, we must be able to address the wide and diverse range of AI workloads across different markets. This is accompanied by extensive computational requirements, meaning we need to offer more than one computing solution, each optimized for the needs of specific markets. As the industry's demand for customized chips continues to rise, coupled with increasing costs and complexity in chip production, Chiplets are gradually becoming a widely adopted solution in the industry."

By reusing specialized Chiplets to develop various customized System-on-Chips (SoCs), higher-performance and lower-power-consumption system designs can be achieved compared to traditional monolithic chips, with lower overall design costs. However, without unified industry standards and frameworks, differences between various chipsets may lead to compatibility issues, thereby hindering the pace of innovation. To address this fragmentation issue, Arm launched the CSA last year. The CSA provides a set of standards for system partitioning and Chiplet interconnectivity developed jointly with the ecosystem, enabling the industry to reach a consensus on fundamental choices for building Chiplets. Through the CSA, new Chiplet designs can seamlessly fit and be reused in compliant systems, which not only accelerates system innovation based on Chiplets but also effectively reduces the risk of fragmentation.

The broad participation of these innovative technology companies in the CSA forms the cornerstone of the Chiplet ecosystem based on Arm technology. This ecosystem is dedicated to revolutionizing system design, making SoCs more flexible, accessible, and cost-effective while significantly reducing the risk of fragmentation. With the publication of open specifications, designers can reach a consensus on how to define and connect Chiplets to build composable SoCs. These SoCs, with high flexibility, can meet the diverse needs of AI workloads and ensure that the final chip products precisely fit the needs of specific markets.

Multiple partners participating in the CSA are also actively building solutions within the Arm Total Design ecosystem project. Arm Total Design is an ecosystem dedicated to seamlessly delivering custom chips powered by Arm Neoverse Compute Subsystem (CSS). To date, Arm Total Design has achieved significant accomplishments in deploying Chiplet-based CSS solutions that support strategic implementations for specific markets, including:

• Customizing AI workloads for diverse markets: Alphawave Semi's customers have urgent needs for high-performance chips for AI workloads across networking, edge computing, storage, and security. In response to these needs, Alphawave Semi combines Chiplets based on Arm Neoverse CSS with proprietary I/O dies, utilizing AMBA CHI C2C technology to interconnect accelerators customized for different market needs. These market-specific customized chips, based on a standardized foundation, can effectively share the cost of compute dies while maintaining the flexibility to build multiple systems.
• Revolutionizing large-scale AI training and inference workloads: ADTechnology, Samsung Foundry, Rebellions, and Arm have jointly created an AI CPU Chiplet platform for training and inference of large-scale AI workloads in data centers, expected to deliver a 2-3x energy efficiency advantage for generative AI workloads (e.g., Llama3.1 with 405B parameters LLMs). This multi-vendor Chiplet platform integrates Rebellions' REBEL AI accelerators, coherent NPUs using AMBA CHI C2C interconnect technology, and ADTechnology's compute Chiplets based on Neoverse CSS V3, and is manufactured using Samsung Foundry's 2nm Gate-All-Around (GAA) process technology. This achievement benefits from the standardization progress made by the CSA.

The CSA has provided efficient solutions for diversified AI-driven workloads across multiple market sectors such as infrastructure, automotive, and consumer electronics, and has established numerous successful cases. Leveraging the exceptional flexibility of the Arm computing platform, seamless communication technologies enabled by standards such as AMBA CHI C2C, and the integration innovation trend led by the CSA, the Chiplet ecosystem based on Arm technology can cleverly address the growing AI demands in various markets. As the CSA ecosystem continues to grow and the industry deepens its focus on standardization and collaboration, Arm will work with partners to significantly reduce fragmentation and accelerate the development and deployment of customized chip solutions.