At a time when China's tech companies are frantically scrambling for Nvidia's high-end GPU chips to catch generative AI business opportunities, Wei Shaojun, a professor at Tsinghua University and an academician at the International Eurasian Academy of Sciences, emphasized at Intel's 2023 China Academic Summit that future AI technology needs to start from the end application -- define software by application first and then define the chip by software.
Wei Shaojun, as quoted by Chinese media China Electronic News, proposed a new chip architecture that the semiconductor industry in China is pursuing today, which not only requires strong software programmability but also strong hardware programmability.
He believes that as the scale of computing power continues to expand and the demand for computing power continues to rise, the new challenge for AI development is how to more efficiently allocate, share, schedule, and release more computing power.
This challenge is especially urgent for the Chinese industry, which is facing the US government's clampdown on the supply of AI chips, especially when there is no lack of concepts and actions to advocate the aggregation of computing power in China, said Wei. He also analyzed the shortcomings of different currently available architectures of chips used for AI applications.
1. CPU, GPU, and other processors, which have strong flexibility in software programmability, but are weaker in hardware.
2. ASICs and SoCs, which are weak in software programmability and the hardware cannot be changed once the chip is completed.
3. Programmable logic devices represented by FPGAs and EPLDs (Erasable Programmable Logic Devices), which are highly flexible in terms of hardware programmability, but weak in terms of software programmability, as well as costly and expensive.
Wei pointed out that the core of implementing AI lies in the software, and the foundation of supporting AI is the chip. Under such circumstances, Wei pointed out that AI still faces two practical problems: one is the continuous evolution of algorithms, as there will always be new algorithms. Secondly, since there is no unified algorithm, a specific algorithm goes with a specific application.
In this regard, he proposes that deep learning is in need of an intelligent computing processor that can be programmable for various applications and can realize the ability to migrate data from the cloud to the edge. Wei said the aforementioned new chip architecture can solve the problems.
From the view of resource limitations, the concept of the aforementioned new architecture actually stems from China's difficulty to obtain AI chips. Wei Shaojun was expressing an urgency for China to develop a new architecture due to current limitations.
However, how to implement that design and from where to start remains to be seen.