The application of AI drives the development of in-memory computing

 Since the development of applications, the emergence of AI has driven the development of computational storage/integrated storage and computing/in-memory computing. The application of storage and calculation integration technology in AI includes Voice cmds,Voice enhancement,Health monitoring,Environment identification,Gesture controls,Sports classification,Object detection,Positioning, etc.

The memory-intensive (big data demand) and computing-intensive (low-precision regular operation) characteristics of AI algorithms provide powerful conditions for the realization of computational storage/integration of storage and computing/in-memory computing:

1) The AI algorithm is a very large and complex network, which contains a large amount of image data and weight parameters, and a large amount of data will be generated during the calculation process. Taking the VGG-16 network as an example, the number of weights is about 1.4*108, processing a 3-channel image with a size of 224*224 requires about 1.5*1010 multiplication and addition operations, and the data needs to be stored in the computing unit and Frequent movement between units. Memory access bandwidth has become one of the important bottlenecks of convolutional neural networks. At the same time, the energy consumption required to move data from DRAM storage to computing units is 200 times that of computing itself, seriously affecting the computational energy efficiency of convolutional neural networks. Therefore, for convolutional neural networks, there is an urgent need for suitable means to reduce data movement and its resulting performance and power consumption overhead.

2) There are a large number of regular operations in the training and reasoning process of AI algorithms, such as multiplication and addition operations. In order to complete a large number of calculations, the general chip design idea is to add a large number of parallel computing units, such as thousands of convolution units. However, with the increase of computing units, the memory bandwidth and capacity that each computing unit can use are gradually decreasing, and memory access has become the performance bottleneck of AI. The core problem encountered in the chip implementation of AI algorithms has shifted from the original strong demand for computing power to the constraints on memory access bandwidth and power consumption. For example, a GPU that can provide a large number of computing resources, its actual computing power and computing resources are greatly reduced during runtime, and computing efficiency is greatly limited by memory access. At the same time, the calculations in AI algorithms do not require high precision.

Therefore, in view of the memory-intensive nature of AI algorithms, as well as the regular operations and low-precision requirements in intensive computing, many researchers have begun to pay attention to computational storage/integrated storage and computing/in-memory computing. And the low-precision computing requirements also provide more possibilities for its realization.