A major breakthrough, Tsinghua University's optical chip can support general AI computing power, which is expected to achieve corner overtaking

Can the United States jam China in the field of artificial intelligence chips? After 920, Tsinghua also took action, and it was Wang Bang who came up.

Miao Wei, former minister of the Ministry of Industry and Information Technology, said that the mainland's computing power is constantly being increased by the United States because of the restrictions imposed by the United States on advanced AI chips, so we should not blindly pursue the hot spots of large computing power, large models, and big data, but to empower various specific fields to realize the application of artificial intelligence and achieve a differentiated development path.

This strategy is far-sighted and reflects the unique wisdom of our Chinese.

Now let's look at where the AI chips in the United States are used? ChatGPT, Sora, etc., most of them are such large models, Microsoft, Google, etc., most of them are such Internet companies.

And we are walking on two legs, one is to develop large models according to our own actual conditions, and the other is more important, focusing on the development of artificial intelligence applications, especially in the field of industrial production and manufacturing, for upgrading.

Of course, the reason why these goals can be achieved is that we can already achieve a certain degree of domestic substitution in the field of AI chips. In fact, when it comes to artificial intelligence chips, there are many domestic manufacturers, all of which have certain technical strength, and they have also achieved a good development trend over the years.

For example, Huazi's 910B can already compete with the A100, and is twice as strong as the H20 that can be purchased now. As for the recently rumored 920, we also talked about it in the previous video, in terms of parameter comparison, it can do back and forth with H100 or even B200, and in terms of energy efficiency, it has achieved a comprehensive surpass. Therefore, this chip is extremely advanced, which can achieve the same computing power and save more power.

So this can actually show two things, one is that in terms of chip design, we must have no problem, we can reach the international advanced level, not to mention Huazi, which can reach the top level. Now the poor is just manufacturing. Therefore, with the emergence of 9000S, substantial breakthroughs have been made in the production and manufacturing process, but it may be temporarily self-sufficient at the moment, and it will take time to achieve a large number of foundries, but this is only a matter of time.

In another 1-2 years, the domestic semiconductor industry chain may achieve a comprehensive breakthrough, and by that time, all domestic AI chip manufacturers will no longer have to worry about manufacturing problems, and they will be able to fully blossom. At the same time, American AI chips may not have any sales in China. After all, if you buy and use it, what will you do if you don't sell it one day? To do computing power, it is not simply to buy a few chips, there are various software systems, development tools, etc., and it is difficult to temporarily replace the chips of other manufacturers.

Then let me tell you another good news, that is, Tsinghua University has developed an optical chip that can be used for general artificial intelligence computing power.

The research group of Associate Professor Fang Lu of the Department of Electronic Engineering of Tsinghua University and the research group of Academician Dai Qionghai of the Department of Automation of Tsinghua University pioneered the distributed breadth intelligent optical computing architecture and developed the world's first large-scale interference diffraction heterogeneous integrated chip Taichi, which can achieve 160 TOPS/W general intelligent computing.

The research results were recently published in the latest issue of Science with the title of "Large-scale Optical Chip "Taiji" Empowers 160 TOPS/W General Artificial Intelligence", which is undoubtedly authoritative.

Inspired by Zhou Yizhong's "Yi You Tai Chi, Shi Shengliang Yi", the scientific research team established an interfero-diffraction joint propagation model, integrated the advantages of large-scale parallelism of diffraction light computing and the flexible reconstruction characteristics of interferometric light computing, and reconstructed and reused the diffraction codec and interference feature calculation partially/whole, broke through the flux bottleneck with time series multiplexing, supported the distributed breadth optical computing architecture from the bottom up, and explored a new path for large-scale general intelligent optical computing on the chip.

To put it simply, interference-diffraction is like a concave building block and a convex building block spliced together to form a base block, and then reconstructed and reused to realize infinite possibilities.

It may still be difficult to understand, it doesn't matter, we just need to know that this "Taiji" optical chip has an energy efficiency of 160 TOPS/W. So how good is this energy efficiency?

The energy efficiency of 920 is 5.714TOPS/W, H100 is 2.872TOPS/W, and the latest B200 is 4.5TOPS/W. Obviously, they can't be compared with 160TOPS/W, and the "Taiji" optical chip is dozens of times higher. It's terrifying, and this may be the power of light chips.

With the continuous development of Moore's Law, the traditional semiconductor manufacturing process has approached its limit, and electronic chips have encountered difficult bottlenecks in further improving computing speed and reducing power consumption.

In contrast, because optical chips are calculated through the information changes in the propagation and interaction of light, they have significant advantages such as fast speed, low latency, and low energy consumption that cannot be matched by silicon-based chips. This time, the "Taiji" optical chip has made a huge breakthrough and realized for the first time the processing of complex tasks of artificial intelligence such as the recognition of thousands of objects in natural scenes and cross-modal content generation. It can be seen that the Tai Chi optical chip will have the opportunity to provide computing power support for large model training and reasoning, general artificial intelligence, and autonomous intelligent unmanned systems in the mainland.

In addition, it should be noted that this is only the first generation of chips, and with the iteration and upgrading of technology, it will definitely be able to obtain more powerful performance and a wider range of application scenarios, which is really worth looking forward to.

In short, the advent of the "Taiji" optical chip means that the mainland has once again made a major breakthrough in the field of AI chips and achieved an epoch-making corner overtaking. Maybe there will be an opportunity in the future, relying on optical chips, to solve the fundamental problem of the mainland's computing power base once and for all.

It is impossible for the United States to kill China in AI chips and completely curb the development of Chinese artificial intelligence. Restricting advanced AI chips can only affect the development speed of some Chinese large models, but unfortunately this is not the most important. And American chip companies will not only lose real money orders, but also lose the future Chinese AI chip market. As far as we are concerned, as long as China's industrialization system can be upgraded and iterative, it doesn't matter what the United States and the West do, it is all in vain, and their goals will never succeed.