When it comes to in memory computing, the first impression of most people is ultra-low power consumption and large computing power. In memory computing technology has broken through the limitations of von Neumann architecture, broken through the memory wall, and brought new innovations to the semiconductor industry. But you may ask, what are the application scenarios of in memory computing?



Next step of edge calculation



Edge computing can be said to be the first step taken by many in memory computing technologies and companies. With its low-power characteristics, wearable and other end-to-end devices can be said to be customized for this technology. Under the advantages of its architecture, in memory computing has more considerable computing power than many traditional edge AI chips. Therefore, for applications such as smart watches and smart glasses that require high power consumption and certain AI computing needs, in memory computing chips are undoubtedly the best choice.



However, today’s MCU has reduced the power consumption to a very low level, and some of them can also complete some simple AI operations. If they only compete in applications such as speech recognition and event detection, even if these in memory calculations have advantages, in the actual use process, in addition to the endurance, the changes perceived by users will be relatively small.

Wtm2101 in memory computing chip / Zhicun Technology


But edge computing is not only limited to this, but also the old difficulty of image processing needs to be solved. This application has higher computing power requirements than those above. Zhicun technology, a leading domestic in memory computing company, recently revealed that they are building a next-generation in memory computing chip with stronger computing power, which is oriented to ultra clear video processing. According to its demonstration, the chip is mainly aimed at Ai framing, AI super-resolution, AI video noise reduction and AI high dynamic resolution, which are AI applications with more obvious perception at the edge.



The above AI applications are only the next step for in memory computing to go to the consumer edge end such as smart phones. The ultimate goal of edge AI chips is automatic driving. If we can achieve higher computing power, the in memory computing chip will have the opportunity to rush into the automobile market and compete with the players of the autonomous driving chip.



Computing power beyond GPU



Since in memory computing has proved its computing power advantage at the edge, is there a chance to compete with GPU and other computing Raptors? We take wave simulation as an example. Wave simulation is popular in many applications, such as medical imaging, oil exploration, earthquake disaster mitigation and national defense systems. However, when most applications use wave simulation, supercomputers are used to solve multiple wave equations. Although such applications do not require high cost and volume like wearable applications, they still attach great importance to speed and energy consumption.



At present, CPU and GPU are the dominant wave simulation solutions. However, due to the lack of parallelism, even the small problem of high-end CPU operation requires a lot of time to complete the calculation. GPU, with its huge memory bandwidth advantage, undoubtedly has a higher speed. Even so, in practical applications, fluctuation simulation is an extreme data movement process, and GPU will still encounter bottlenecks. Even if the memory bandwidth of hundreds of gb/s can not be affected, the power consumption for data movement is even higher than that for computing.



In memory computing can reduce the data movement between processors, because it eliminates the data movement between off chip and on-chip storage, but the data movement between in memory is still a big problem. ExxonMobil researchers came up with wave PIM, an in memory computing scheme, which uses the h-tree architecture commonly used in VLSI to reduce the delay of data movement between memory blocks. They simulated with 16GB hbm2 memory with 900gb/s bandwidth, and obtained the result of 52.8tflops (fp32), which exceeded Tesla V100 GPU. This proves that memory computing chips have unique advantages even in server level and HPC level applications.

UPMEM PIM / UPMEM


However, the GPU memory bandwidth has reached 3tb/s with hbm3 and NVIDIA’s H100 chips, and the DDR4 PIM of French company upmem, which is the industry’s dominant in memory bandwidth, has only achieved 2.5tb/s. Even though in memory computing has a huge advantage in power consumption, it still needs more advanced memory technology and more architecture innovation to further surpass GPU in performance. Fortunately, now more and more companies are trying to commercialize in memory computing. Although the storage manufacturers have not decided to go in this direction, there is no conflict between in memory computing and its development technology. Moreover, from the perspective of production innovation and investment, they have begun to lay out this technology. In the future, it is likely that storage manufacturers and GPU manufacturers will be mutually involved in high-performance computing.

Leave a Reply

Your email address will not be published. Required fields are marked *