ByteDance, the Chinese tech company that owns TikTok, is significantly changing its approach to AI hardware by using Huawei chips to train its latest AI model.
While doing this, ByteDance has also maintained a substantial presence in the NVIDIA ecosystem. It continues to be a major client for NVIDIA products which are accessible via cloud computing through Microsoft. Last year, the company allocated approximately $2 billion for NVIDIA’s H20 AI chip, tailored for the China market in response to trade restrictions.
This development comes against the backdrop of China rapidly closing the gap in AI capabilities, now only 6-9 months behind the US. A few months ago, 01.AI chief Kai-Fu Lee said, “We were six or seven years behind.” But now, China seems to be ahead of the US.
Recently, Chinese researchers introduced the Diagram of Thought (DoT), a novel framework that models reasoning as a directed acyclic graph (DAG). This surpasses OpenAI’s o1 by enabling non-linear iterative thinking, better reflecting human cognitive processes, and improved complex reasoning in LLMs.
Case in point: As ByteDance navigates this complex landscape, Huawei’s upcoming third-generation AI processor chip, Ascend 910C, could significantly impact the Chinese market.
Huawei, a global provider of information and communications technology (ICT) infrastructure and smart devices in China, stands to provide a competitive alternative to NVIDIA’s offering of its B20 chips with its 910C chips, expected to be released in October 2024.
NVIDIA Struggles
With NVIDIA’s introduction of its Blackwell GPU architecture early this year, it was reported to have control over 95% of the AI chip market, making it larger than Tesla and Amazon combined. As reported by AIM, several non-NVIDIA companies have recently risen with their AI-powered processing chips.
While NVIDIA’s GPUs are in great demand because they are instrumental in training AI models, they might not be the best tool available for inference.
When approached for a response by AIM, NVIDIA declined to provide any comments on the matter.
Former Investment Bank Trader Pablo Heman says on X, “The troubling sign is that NVIDIA gross margins fell from 78% to 75%. It shows me that NVIDIA’s pricing power is not as strong as before.”
I understand Nvidia just saved America again, but the threats are coming thick and fast! First AMD has it new GPU to release soon and then Huawei has its Ascend 910C to launch in October. The troubling sign is that NVDA gross margins actually fell from 78% to 75%, it shows me… pic.twitter.com/twRMxriIrk
— Pablo Heman (Plan A) (@RealPabloHeman) August 29, 2024
ByteDance’s Shift to Huawei
ByteDance has reportedly ordered over 100,000 Ascend 910B chips but has received fewer than 30,000 as of July 2024, highlighting supply chain challenges that could impact its AI development timeline.
The Ascend 910B chips are used for less computationally intensive tasks, such as operations where pre-trained models make predictions. However, training new AI models requires significantly more computational power, typically provided by high-performance GPUs like those from NVIDIA.
This shift towards Huawei’s chips may limit ByteDance’s ability to develop more complex AI models compared to its existing offerings, such as the Doubao chatbot, which has gained popularity among Chinese users.
Despite its pivot towards Huawei, ByteDance remains one of the largest buyers of NVIDIA chips tailored for the Chinese market.
Huawei’s Ascend 910C Chips
Huawei’s upcoming Ascend 910C chips are a game-changer in the Chinese market. It could be successful if ByteDance integrates them into its AI infrastructure upon its release, likely in October 2024.
The Ascend series has been developed as a competitive alternative to NVIDIA’s offerings. It aims to meet China’s growing demand for AI capabilities amid geopolitical tensions and trade restrictions.
The Chinese government has been actively encouraging technology firms to source AI components domestically. This push not only aims to bolster local manufacturers but also seeks to reduce dependency on foreign technology providers like Intel, AMD and NVIDIA.
Regardless, many still doubt such changes will happen due to speculations about the high transfer costs and low manufacturing of domestic products.
The limited supply of Ascend 910B chips poses a significant obstacle for ByteDance as it attempts to scale its AI initiatives. The slow pace of chip delivery could delay development timelines and hinder the company’s ability to compete effectively in the fast-evolving AI landscape.
Additionally, there are concerns about Huawei’s chips’ overall performance compared to established players like NVIDIA.
Tony Peng, the global head of communications at Baidu said, “Even though Huawei’s 910C can outperform B20 and catch up with H100, its gap against Nvidia in terms of software ecosystem (CUDA) and chip-to-chip will affect the performance when tens of thousands of chips are connected.”
As Huawei prepares to launch its Ascend 910C chips, the success of these products will be closely watched by ByteDance and other Chinese tech firms.
The future of AI development in China hinges on how effectively companies can leverage local resources while balancing their import needs.
The post ByteDance Turns to Huawei Amid US Export Restrictions appeared first on AIM.
