칩렛의 표준화를 주도하는 Arm이 CSA 첫 번째 공개 사양을 발표했다. 업계는 파편화를 대폭 줄이고 맞춤형 실리콘 솔루션의 개발 및 배포를 더욱 가속화하게 될 것으로 기대했다.
CSA Disclosure, Partner Case Presentation
Arm, which is leading the standardization of chiplets, has released the first public specification of the CSA, which the industry expects will significantly reduce fragmentation and further accelerate the development and deployment of custom silicon solutions.
Arm announced on the 3rd that it has presented the first public specification of the Chiplet System Architecture (CSA).
Reusing specialized chiplets to create multiple custom systems-on-chips (SoCs) enables the creation of systems with high performance and low power consumption while reducing design costs compared to monolithic chips.
However, chiplets can cause compatibility issues due to the diversity of chipsets without industry-wide standards and frameworks, and to address this fragmentation, Arm introduced the Chiplet System Architecture (CSA) last year.
CSA provides a set of system partitioning and chiplet connectivity standards, co-developed with the ecosystem, to enable the industry to tailor the basic choices for building chiplets.
Currently, more than 60 leading companies, including AD Technology, AlphaWave Semi, AMI, Cadence, Jaguar Micro, Kalray, Rebellion, Siemens, and Synopsys, are participating in CSA to contribute to silicon strategies and implement standards in various market segments.
Several suppliers participating in CSA have already built solutions as part of Arm Total Design, an ecosystem that provides custom silicon powered by the Arm Neoverse compute subsystem (CSS). To date, Arm Total Design has successfully deployed chiplet-based compute subsystems.
Alphawave Semi combines Arm Neoverse CSS-based chiplets with a proprietary I/O die, allowing Alphawave to connect accelerators tailored to the specific needs of each market using AMBA CHI C2C. Tailored SKUs for specific markets are derived from the standard base, reducing the cost of the compute die while maintaining the flexibility to build multiple systems.
ADTechnology, Samsung Foundry, Rebellion, and Arm have combined technologies to develop an AI CPU chiplet platform for training and inference of large-scale AI workloads in the data center.
Llama3.1 405B parameter LLM is expected to have an efficiency advantage of about 2-3x, and the multi-vendor chiplet platform combines Rebellion's REBEL AI accelerator and a coherent NPU using the AMBA CHI C2C interconnect.
As a result of the standardization work with CSA, it will be built on ADTechnology’s Neoverse CSS V3-based computing chiplet that can be implemented with Samsung Foundry’s 2nm GAA advanced process technology.
“AI has the potential to permeate every market and unleash a new industrial revolution, requiring the ability to address a wide range of AI workloads across diverse markets,” said Eddie Ramirez, vice president and general manager of the Infrastructure Business Unit at Arm. “This means offering more than one compute solution, optimized for specific market needs. The increasing demand for custom silicon, combined with the cost and complexity of producing silicon, is driving the adoption of chiplets.”