ByteDance, the parent company of TikTok to collaborate with TSMC to develop its own AI GPUs. According to reports, the two AI GPUs will enter mass production by 2026. By producing its own chips company aims to reduce its dependency on Nvidia’s pricy GPUs.
According to reports, the parent company of TikTok, ByteDance is developing its own AI GPUs with assistance from the Taiwan Semi-Conductor Manufacturing Company (TSMC) thereby aiming to reduce reliance on Nvidia for AI hardware.
The company has already designed two chips and is planning mass production with TSMC by 2026. The self-designed chips could save billions of dollars compared to buying Nvidia AI GPUs.
ByteDance AI chips are supposedly to use TSMC’s 5nm node, one generation behind the most advanced version. The two AI GPUs are still in the design phase, including one for the AI interface and another for AI training. The generated chips are very similar to Nvidia’s next-generation AI chip, Blackwell.
The increased spending on the generative AI model compelled ByteDance to produce their own chips. The information estimates that ByteDance spent over 2 billion US dollars for buying over 200,000 chips earlier this year. Many of these GPUs are yet to be delivered to the company.
In response to the export controls made last year, Nvidia has launched AI chips H20, L20, and L2 to attract the Chinese market.
The H20 GPU has a strong demand for AI servers from Chinese cloud service providers including Tenant and Huawei, despite the fact that the H20 GPU has lower performance and fewer cores than the H100. According to reports Nvidia’s processor beats Huawei’s contenders.
Training AI models are computationally demanding, making Nvidia GPUs well known. The strategic inclination towards self-production is subject to high challenges.
According to experts, ByteDance’s current AI models such as Doubao and Jimeng were designed using powerful hardware. Shifting to TSMC’s chips may obstruct training complex models.
Even though ByteDance’s shift to self-production was mainly to reduce cost and increase on-time delivery, the chances of making GPUs faster than Nvidia HGX H20 are unlikely. This is due to the underlying United States export rules making TSMC unable to ship GPUs to ByteDance, which is located in China.
Another risk factor regarding self-production is the need to set up their own software platform that must be compatible with the hardware solutions. Though many Chinese companies have developed AI GPUs they are for specific tasks only.
The US set of controls released earlier this month foresees an increased effort to hinter China’s technological advancements making the ability of Chinese companies to work with TSMC regulated under control.
ByteDance decision for self-bulk production seems to navigate US trade restrictions while searching for local operators for help to maintain its roots in AI. The recent investment of the company in Xinyuan Semiconductors, a Chinese chip manufacturer to develop a VR headset showcases its potential to compete with Meta’s Quest and Apple’s Vision Pro. The new venture is quite formidable, however, if succeeds, it may be an inspiration for many other companies.
Bottom line
ByteDance decision to mass-produce AI GPUs includes numerous factors, the existing US export control rules, high-priced Nvidia AI GPUs, increasing demand, and lack of on-time delivery. Nvidia is a well-qualified company and producing a chip similar to them is quite demanding.
Many companies have launched their own chips but for specific purposes, others quietly depend on Nvidia. However, the self-production saves billions of dollars and also paves a new path for conquering more roots in AI and beyond if succeeds.
Latest Tech News: New Google Play Store Warning: 200 Dangerous Apps with 8 Million Downloads Exposed
