Niphon | Istock | Getty Pictures
Tencent and Baidu, two of China’s largest know-how corporations, revealed how they’re conserving within the international synthetic intelligence race even because the U.S. tightens some curbs on key semiconductors.
The enterprise’ strategies embody stockpiling chips, making AI fashions extra environment friendly and even utilizing homegrown semiconductors.
Whereas the administration of U.S. President Donald Trump scrapped one controversial Biden-era chip rule, it nonetheless tightened exports of some semiconductors from corporations together with Nvidia and AMD in April.
Large names within the sector addressed the difficulty throughout their newest earnings convention calls.
Martin Lau, president of Tencent — the operator of China’s greatest messaging app WeChat — stated his firm has a “fairly sturdy stockpile” of chips that it has beforehand bought. He was referring to graphics processing models (GPUs), a sort of semiconductor that has turn into the gold commonplace for coaching enormous AI fashions.
These fashions require highly effective computing energy equipped by GPUs to course of excessive volumes of knowledge.
However, Lau stated, opposite to American corporations’ perception that GPU clusters have to develop to create extra superior AI, Tencent is ready to obtain good coaching outcomes with a smaller group of such chips.
“That truly kind of helped us to take a look at our current stock of high-end chips and say, we must always have sufficient high-end chips to proceed our coaching of fashions for a number of extra generations going ahead,” Lau stated.
Relating to inferencing — the method of truly finishing up an AI activity quite than simply coaching — Lau stated Tencent is utilizing “software program optimization” to enhance effectivity, with a view to deploy the identical quantity of GPUs to execute a selected perform.
Lau added the corporate can also be wanting into utilizing smaller fashions that do not require such giant computing energy. Tencent additionally stated it could possibly make use of custom-designed chips and semiconductors presently obtainable in China.
“I believe there are numerous methods [in] which we will fulfill the increasing and rising inference wants, and we simply have to kind of maintain exploring these venues and spend most likely extra time on the software program aspect, quite than simply brute pressure shopping for GPUs,” Lau stated.
Baidu’s strategy
Baidu, China’s greatest search firm, touted what it calls its “full-stack” capabilities — the mixture of its cloud computing infrastructure, AI fashions and the precise functions based mostly on these fashions, resembling its ERNIE chatbot.
“Even with out entry to essentially the most superior chips, our distinctive full stack AI capabilities allow us to construct sturdy functions and ship significant worth,” Dou Shen, president of Baidu’s AI cloud enterprise, stated on the corporate’s earnings name this week.
Baidu additionally touted software program optimization and the power to deliver down the price of working its fashions, as a result of it owns a lot of the know-how in that stack. Baidu administration additionally spoke about efficiencies that permit it to get extra out of the GPUs it possesses.
“With basis fashions driving up the necessity for a large computing energy, the skills to construct and handle giant scale GPU clusters and to make the most of GPUs successfully has turn into key aggressive benefits,” Shen stated.
The Baidu government additionally touted the progress made by home Chinese language know-how corporations in AI semiconductors, a transfer he stated would assist mitigate the influence of U.S. chip curbs.
“Domestically developed self-sufficient chips, together with [an] more and more environment friendly home-grown software program stack, will collectively kind a robust basis for long-term innovation in China’s AI ecosystem,” Shen stated.
China home chip focus
China has been ramping up improvement of chips designed and manufactured on its residence soil for the previous couple of years. Most consultants agree that Beijing stays total behind the U.S. within the realm of GPUs and AI chips, however there have been some advances.
Gaurav Gupta, an analyst masking semiconductors at Gartner, stated stockpiling is a technique Chinese language corporations are coping with export restrictions. Moreover, there was some progress made in semiconductor know-how in China, even when it stays behind the U.S., Gupta added.
“China has additionally been creating its personal home semiconductor ecosystem, all the best way from supplies to tools to chips and packaging. Completely different segments have made various ranges of progress, however China has been surprisingly extraordinarily constant and bold on this aim, and one should admit that they’ve achieved first rate success,” Gupta instructed CNBC by electronic mail.
“This gives an avenue for them to obtain AI chips, which maybe cannot compete with these from the united stateschip leaders however proceed to make progress.”
Many U.S. executives have urged Washington to scrap export restrictions in mild of China’s progress. Nvidia CEO Jensen Huang known as the curbs a “failure” this week, saying they’re doing extra harm to American companies than to China.
