US commerce restrictions have created important obstacles for Chinese language corporations, limiting their entry to superior AI {hardware} wanted to stay aggressive globally.
Nvidia’s H20 GPUs, scaled-down variations of the highly effective H100, have been developed to fulfill export management necessities however nonetheless include a hefty price ticket of round $10,000 per unit.
Even at that worth, the provision of those GPUs is restricted, compounding the difficulties Chinese language corporations face. This scarcity has fueled a thriving black marketplace for Nvidia’s high-end chips, such because the H100 and A100, the place costs proceed to rise resulting from overwhelming demand. Nonetheless, world corporations, notably ByteDance – the mum or dad firm of TikTok, already beneath intense scrutiny within the US – can’t afford the authorized and reputational dangers related to participating in such illicit markets.
Two AI chips
ByteDance has made important investments in AI, reportedly spending over $2 billion on Nvidia’s H20 GPUs in 2024, and now in keeping with The Info, the corporate is trying to develop its personal AI GPUs to scale back dependency on Nvidia.
The report provides these chips will embody one designed for AI coaching and one other for AI inference, and each can be produced utilizing TSMC’s superior N4/N5 course of, the identical expertise used for Nvidia’s Blackwell GPUs.
Broadcom, acknowledged for its AI chip designs for Google, will reportedly lead the event of those GPUs, that are anticipated to enter mass manufacturing by 2026. Whereas a number of Chinese language corporations have developed their very own AI GPUs to scale back reliance on Nvidia, most nonetheless rely on Nvidia’s {hardware} for extra demanding duties. Whether or not ByteDance can totally transition to its personal {hardware} – and whether or not it might wish to – stays to be seen.
The transfer will definitely not be with out challenges. As Tom’s {Hardware} notes, “The corporate now depends on Nvidia’s CUDA and supporting software program stack for AI coaching and inference. As soon as it goes with its AI GPUs, it should develop its software program platform and guarantee its software program stack is totally appropriate with its {hardware}.”
Extra from TechRadar Professional
Source link