SAL’s research focus is a novel computer architecture for future AI computing and Cloud/IoT platforms. Its topics include Distributed Hardware System, GP-GPU, Neural Processing Unit, Memory sub-system, and Reliability. These are the critical aspects to scale up/out existing computers to deal with massive data in the era of Artificial Intelligence and Machine Learning.

As processing scaling continues, transistors became more vulnerable to faults and errors. In memories, there are new types of errors, such as variable retention time and row hammer. In logics, increasing variability makes it more difficult to exploit faster transistors from the scaling: with fewer atoms to build a transistor, few more atoms now greatly affect the performance of the transistor. Because of this process variation, there has to be more timing margins to run safe.

AI and Big Data made Memory Sub-system crucial in throughput-oriented processors, like NPU and GPU. These processors has increased computation throughput over decades, but the amount of data to feed these computation units has grown relatively poorly. To improve system performance, balancing memory storage capacity and transfer bandwidth with computation throughput is one of the key challenges in modern processors.

Our base platform to compute is migrating fast from the PC/mobile + server model to the CLOUD/EDGE/IoT platform, as the amount of data keeps exploding. This new platform partitions data and its computations across cloud, edge, and sensor layers, based on the amounts of resources available and resources needed. With the new platform and partitioning, network efficiency will be the key factor to integrate these layers and will determine overall performance of the platform.