Can chiplets power China’s genAI ambition?

Baidu, ByteDance, Tencent, and Alibaba have collectively spent US$1 billion to acquire approximately 100,000 A800 processors from US chipmaker Nvidia.

Additionally, these firms have purchased another US$4 billion in graphics processing units (GPUs) set to be delivered by 2024.

This comes just before the US ban on exports of Nvidia’s A100 and H100 chips.

Nvidia office / Source: Nvidia

That’s why Chinese companies are now exploring alternative options that may not rely on traditional GPUs.

One such option is chiplets. Chinese authorities are increasingly emphasizing the importance of the tech, as per a Reuters report. There has also been a growing number of policy documents emerging from China that mention chiplets, and Chinese companies are investing in or acquiring firms working on these units.

A chiplet is a small, individual semiconductor die or module that is used as a building block to construct a larger and more complex integrated circuit.

Microelectronics engineer working in a lab/ Photo credit: Shutterstock

Instead of designing and manufacturing a single chip to perform all functions, chiplets can be combined or stacked to carry out a similar purpose.

Chinese companies like Chipuller are trying to beat US chipmakers by manufacturing these units. Recently, the company acquired 28 patents – either owned by Silicon Valley-based firm zGlue or developed by individuals named in zGlue’s patents.

China’s purchase of chip packaging equipment – which can assist in large-scale chiplet manufacturing – is also ramping up, rising to US$3.3 billion in 2021. However, that number fell last year to US$2.3 billion, according to the earlier Reuters report.

Chinese companies craving for chiplets

One of the advantages of chiplet technology is its flexibility, believes Bob Grim, vice president of strategy and corporate communications at Canada-based chips startup Tenstorrent.

“One of the major challenges that Nvidia faces is the incorporation of very expensive memory into their chips, specifically high-bandwidth memory,” he tells Tech in Asia. Grim believes a large portion of Nvidia’s customer base doesn’t necessarily require high-bandwidth memory.

Chiplets a cheaper option?

China may lose chiplets battle too