With the development of cloud computing, big data, artificial intelligence and the increasing capacity demand for memory products, semiconductor memory has also derived new memory product forms such as computational storage, distributed storage, persistent memory based on new memory particles and storage interfaces, etc., which have higher requirements for the architecture and power consumption of memory control chips, multi-channel parallel management capabilities for storage chips, data security, etc., making the value of memory control chips in memory products continue to increase.
(1) Computational storage
Computational storage refers to the use of the computing power of the storage controller chip to hand over part of the data processing tasks originally performed by the host CPU to the storage control chip, reducing the computing burden of the CPU. Computational storage is good at handling data-intensive applications, and in data processing and data acceleration in scenarios such as database management, ** processing, artificial intelligence layer, and virtualization, it has faster response time and lower latency than CPUs, which can save bandwidth, reduce energy consumption, and achieve high security and privacy because data is processed inside the memory.
Computational storage is standardized and promoted by the Global Network Storage Industry Association (SNIA) and is supported and applied by industry leaders such as Samsung Electronics, Intel, NVIDIA, and IBM. In November 2020, Xilinx and Samsung Electronics announced the Samsung SmartSSD Computational Storage Drive (CSD), the industry's first customizable, programmable computational storage platform that pushes computing power into memory to accelerate applications by more than 10x.
(2) Distributed storage
Distributed storage refers to the decentralized storage of data on multiple storage servers and the integration of scattered storage resources into virtual storage devices, effectively improving the storage and reading efficiency of the storage system. With the development of cloud computing, big data, and artificial intelligence, massive data makes distributed storage gradually replace the traditional way of storing all data on a single storage server. The storage system composed of high-density NAND flash memory particles and high-performance memory control chips is the mainstream solution and future evolution direction of distributed storage.
Using the storage control technology and intelligent storage granular management of the high-performance storage control chip, it can realize high-performance, low-latency and complete storage functions, and provide effective data deduplication, compression and stable data protection functions, so as to meet the new application scenarios with large amount of unstructured data, high data value and diversified storage workload characteristics such as big data analysis, advanced monitoring, high-performance computing, industrial Internet, medical imaging, and virtualized cloud computing. Huawei's full range of distributed storage can achieve 91The highest hard disk space utilization rate of 6%, stable latency of less than 1ms, support for 4,096 node expansion, meet the needs of diverse data analysis services, flexible on-demand purchase and deployment, save space and energy consumption, and reduce TCO (total cost of ownership) by 30%.
(3) Persistent memory based on new storage particles and storage interfaces
Persistent memory is implemented by a new type of memory chip and a new memory controller chip, which adds a layer between memory and SSD to the existing storage hierarchy, providing more storage capacity than memory and lower access latency than SSDs at the same cost. Persistent memory is suitable for application scenarios with storage workload characteristics such as high performance, large capacity, and persistence, and is good at handling tasks such as artificial intelligence, machine Xi, and high-performance computing. The host CPU can perform these types of work tasks directly with direct access to persistent memory, eliminating the need to switch data blocks back and forth between memory and SSD.
This architecture reduces the memory usage requirements that are high at a cost and increases the utilization of SSDs to accommodate different data types, technical needs, and budget constraints. In order to make full use of the characteristics of persistent memory, the CXL protocol, which is built on top of the PCIe logical and physical layers, can standardize the management of persistent memory, and better play the value of persistent memory, has become the mainstream interface development direction. The new CXL-based storage controller chip realizes a variety of new storage protocols defined in the CXL protocol, which can support memory, persistent memory, etc., expand storage capacity, and support resource sharing (memory pooling) and exchange, improve the efficient processing of data and system computing speed, so that persistent memory can be quickly applied to the landing.