Researchers at The George Washington University are developing a flexible interconnection network design, called Adapt-Net, for chiplet-based manycore architectures. The goal of Adapt-Net is to support the concurrent communication of diverse applications running at the same time, improving the energy-efficiency and performance of the manycore architecture. Today’s manycore architectures are often deployed to concurrently execute multiple applications. These applications are mapped into different regions of the manycore system, each of which requires significantly different computing resources (e.g. CPUs, GPUs, and accelerators), memory modules (e.g. caches, and memory controllers), and communication supports (e.g. interconnection networks). However, current interconnection networks optimize the performance for the entire system without considering the diverse demands from different regions of the system.
Adapt-Net can support the diverse communication demands from applications running concurrently. Specifically, any region of the Adapt-Net is capable of adapting to a given topology such as mesh, torus, cmesh, or tree, thus satisfying application's needs in terms of performance and power consumption. For example, as compared to a conventional NoC fabric as shown in Fig. (a), Adapt-Net is able to provide each of the running applications (applications 0-3 shown with different colors in Fig.(b)), with a separate NoC region, that is of different size, location, and topology. As a result, the deployment of Adapt-Net can lead to substantial performance improvement and energy saving for the system. These computing platforms are at the heart of the economic, defense, education, healthcare, and entertainment computing engines of the nation.
Applications:
Chiplet-based parallel computing systems (manycore CPU, GPU, FPGA, TPU, etc.). Manycore processor design for the key players of high-performance parallel computing market, including IBM Corporation, Hewlett Packard Enterprise Company, Intel Corporation, Qualcomm , Inc., Microsoft Corporation, Facebook, Inc., Google, LLC., ARM Ltd, Huawei Technologies Co. Ltd, Cisco Systems, Inc., Advanced Micro Devices, Inc., Micron Technology, Inc, Fujitsu Ltd, Samsung Electronics Co. Ltd, Oracle Corporation, Dell, Inc., Hitachi Ltd (Hitachi Data Systems, Inc.), etc.
Advantages:
Adapt-Net is able to dynamically allocate multiple disjoint topologies at the same time, supporting multi-application execution. Adapt-Net can be dynamically configured to network topologies in response to dynamic application mappings. These topologies can be but not limited to ring, mesh, cmesh, torus, and tree. Full system simulation shows that Adapt-Net can achieve significant latency reduction, energy saving, and execution time reduction compared to prior designs.
