Edge computing, artificial intelligence, machine learning and security are all applications that are important for cloud and networking service providers and can benefit from domain-specific accelerators.
Domain-specific accelerators are silicon-on-chip (SoC) devices that are optimized to deliver cost-effective acceleration of specific applications and networking functions.
The Open Domain-Specific Accelerator (ODSA) Workgroup is an industry initiative driven by seven leading silicon companies to develop an open architecture for domain-specific accelerators. This could dramatically increase the range of opportunities for service providers to accelerate edge computing and related applications.
The bandwidth demands on server processors has significantly outgrown the increases in CPU performance during the past few years. Domain-specific accelerators such as SmartNICs, machine learning/inference coprocessors and security processors can be used to offload processing from CPU cores. This can significantly increase application performance and release additional CPU cores for other revenue-earning workloads or allow the use of less powerful, and significantly lower-cost, processors in edge computing applications. To deliver the best benefits, these accelerators should integrate best-of-breed components including processors, hardware engines, memory and I/O peripherals.
The shift to smaller silicon geometries has enabled larger SoC devices with big caches and many cores but each new silicon technology generation requires huge investment and can result in significantly higher development and die costs. This is becoming a key issue for the development of SoC devices using the 7nm technology that has recently started in volume production. This is causing even the largest semiconductor companies to look at alternatives to a monolithic SoC design for all but the highest-volume applications.
Chiplets are silicon die for subsystem components that can provide an alternative to monolithic SoC design. The chiplets are mounted on a multi-chip module (MCM) which can be configured for different SoC requirements. Chiplets have been widely used where SoC blocks are developed on different silicon technology, for example where you have an analog interface block and a DSP or a processor. Chiplets are also increasingly being used to provide modular solutions. Recent examples using chiplets are the AMD EPYC processors, the Marvell MoChifmodular SOC architecture developed for smartphone applications and high-bandwidth memory (HBM) using 3D stack memory blocks with an interposer.
Most chiplets have, to date, been used in closed development environments where the SoC developer defines all the interfaces and the functionality, and, in many cases, has also developed the chiplets. An open chiplets ecosystem will enable wider use of chiplets and accelerate SoC development, as MCMs are quicker to develop than monolithic SoCs.
The ODSA reference architecture, which is nearing completion, defines common component chiplets and a switching and interface chiplet to which all the other chiplets are connected. The reference architecture will support a communication agent protocol stack to implement inter-die networking. The network infrastructure offers a memory model and performance that is designed to be comparable to a monolithic architecture.
Open domain-specific accelerators and the ODSA reference architecture were discussed during a recent Light Reading webinar that featured speakers from Netronome and Avera Semiconductor (formerly Global Foundries ASIC). In polls conducted during the webinar, 50% of respondents said they were working on domain-specific accelerator solutions now, or would need solutions within one to two years, and almost 30% said their primary application was edge computing. An archive of the webinar -- Open Domain-Specific Accelerators: What Are They & Why Should They Matter to Cloud & Network Providers? -- can be found on the Light Reading website and is free for anyone to access (registration required).
The webinar is based on a white paper produced by Heavy Reading and sponsored by Netronome.
Open specifications are key to a strong chiplets ecosystem and will enable a wider range of domain-specific accelerators for many applications, including edge computing. The ODSA architecture and related specifications will help the development of the chiplets with open interfaces and a common protocol stack that are required to build a strong ecosystem. If you wish to help build this ecosystem, or partner with the ODSA and its members, please contact firstname.lastname@example.org or come to the ODSA Workshop in Santa Clara, CA, on January 29, 2019. For more information, visit: netronome.regfox.com/odsa-workshop.
This article was sponsored by Netronome.
— Simon Stanley, Analyst at Large, Heavy Reading