Opening a new era of computing power
On October 21st, Zhicun Technology How to verify speech recognition based on the in-memory computing chip development board.
01. Compilation tool chain, algorithm...
Ecological finishing point: industrialization of in-memory computing
In the theme sharing session, Zhang Aifei, head of the compilation tool chain of Zhicun Technology, introduced that the WitinMapper compilation and compilation tool chain independently developed by Zhicun maps the algorithm to the WTM2101MPU accelerator through multi-level IR conversion, assisting based on integer linear programming (ILP). The unique matrix block mapping strategy and novel matrix block operation optimization technology not only greatly shorten the development cycle of mapping neural network algorithms to in-memory computing chips, but also improve the computing resource utilization of the storage and computing array and improve the calculation accuracy.
Zhang Aifei, head of compilation tool chain at Zhicun Technology
Algorithms are a key factor in realizing artificial intelligence applications and complement AI chip design and optimization. Yang Xitong, head of algorithm development at Zhicun Technology, shared on the spot: In customer tests, based on market mainstream algorithms, the performance advantage of in-memory computing chips can be several times stronger than traditional computing architecture chips; after optimizing for in-memory computing chips, the algorithm actually runs The delay, power consumption and effect have more obvious advantages. Taking voice command recognition as an example, Zhicun Technology uses deep learning algorithms, and the recognition rate can reach more than 97%. At the same time, the chip operating current is not higher than 0.3mA, and the delay is lower than 50ms.
Yang Xitong, head of algorithm development at Zhicun Technology
02.The first in-memory computing Workshop
Verification of speech recognition based on WTM2101 evaluation board
At this Twin Cities Carnival event, Zhicun Technology and CSDN planned the first in-memory computing workshop: based on the WTMDK2101-X3 evaluation board, taking the provided Chinese and English wake word data sets as an example, the model training was completed and transplanted, and it ran successfully Consider the task completed.
Unlike most workshops, this task integrates hardware and software, including multiple steps such as algorithm training and quantification, algorithm model conversion, algorithm model programming, and algorithm running. The AI enthusiasts present had an in-depth experience of the algorithm training and transplantation process based on in-memory computing architecture chips.
03. Welcome more new forces
Explore the future of AI with us
Interestingly, the first technology enthusiasts to complete the task in Beijing and Hangzhou were all born after 2000, and even most of the participants were emerging forces in the technology field. As one of the award-winning students from the Beijing Railway Station recorded in his blog:
For me personally, I have very little exposure to AI and have only had some experience in the front-end and back-end and embedded fields. Therefore, although this activity may be a relatively easy task for seniors in the AI field, this sense of accomplishment makes me very satisfied, and I can generally understand the role of each process.
...I can gradually feel that AI (especially the current AIGC) has an equally important impact on the embedded field. At this stage, my learning is still limited to some simple controls, such as RTOS. Maybe after deepening some study, you may also learn something in the field of embedded deep learning in the future~
——Reposted from CSDN blogger: Gray Sea Loose
The feedback given to us by every technology enthusiast inspires Zhicun Technology to continue to improve technically and ecologically. Since its establishment, Zhicun Technology has always represented the field of in-memory computing chips and is a pioneering force in new computing architecture and advanced computing power. The in-memory computing training camp will continue to be held, so stay tuned!