Huawei’s attempts to fill the Nvidia-sized gap in China’s AI market are being hindered by performance issues and slower connectivity speeds with its hardware alternative.
Huawei has promoted its Ascend AI chips as a potential alternative to Nvidia's hardware, as the U.S. company is prohibited from exporting its GPUs to China under export controls by the Biden administration.
The FT reports, however, that Chinese AI firms using the Ascend chips have complained over the hardware’s performance.
Subscribe today for free
The Ascend line of chips reportedly suffers from stability issues, a key issue as developers building AI models require hardware to run for extended periods during training efforts.
Users have also reported Ascend suffers from slower inter-chip connectivity.
Inter-chip connectivity enables developers to combine hardware into clusters that power inference and training workloads for AI models. Slower connectivity would mean training processes take far longer.
In addition to hardware troubles, users have reportedly expressed concerns over the related software, dubbed Cann.
Cann is meant to be Huawei’s equivalent of Nvidia’s Cuda software platform which helps developers intuitively during data processing.
However, the FT reports that even Huawei’s staff have expressed concern over Cann’s performance, suggesting it makes Ascend hardware “difficult and unstable to use.”
Another engineer with reported knowledge of Baidu’s use of the Ascend chips told the FT that the hardware crashed frequently.
Huawei had touted its latest AI chip, the Ascend 910B, as being comparable to Nvidia's A100 GPU in terms of computing power and performance.
However, the A100 has since been replaced by the H100, which itself will be usurped by Nvidia’s Blackwell series in 2025, after the company fixed design flaws that had delayed production.
Zhang Ping'an, the CEO of Huawei Cloud Computing Technologies, said in July at the World Artificial Intelligence Conference in Shanghai that China can still be a leader in AI despite restrictions on its access to top-level hardware.
Demand for the company’s Ascend chips to power that leadership role has been strong, with Huawei reporting a 34% increase in revenues in the first half of 2024.
However, the company is reportedly struggling to meet customer demands. Chinese chipmakers face restrictions on importing machinery essential for chip production.
In July, the Korean news site Chosun.biz reported that a low yield rate at Semiconductor Manufacturing International Co. (SMIC), which manufactures the chips for Huawei, impacted production, with a reported four of five processes being defective.
Huawei declined to comment when contacted by Capacity.
Huawei’s AI chipset is “not known for its user friendliness,” according to Lian Jye Su, chief analyst, applied intelligence at Omdia.
“A large part of this is due to the relatively young product maturity - Cann was only introduced in 2018. In comparison, CUDA has been around since the early GPU days. Even the Huawei team themselves are aware of their shortcomings versus Nvidia products,” Su said.
“On the flip side, I also agree that Huawei has excellent client service. They are known to provide responsive and comprehensive on-site support to their clients.
“As such, I don't think this news will dent Huawei's AI ambition. From what I learned about their commercial performance, the Huawei Ascend sales team has long hit their sales target, largely thanks to the enforced decoupling and restrictions on GPU by the US. Most customers simply do not have a better alternative.”
Huawei could soon face competition in China from a familiar face as Nvidia is working on AI chips specifically designed for the Chinese market.
Nvidia has been quietly working on a version of its upcoming Blackwell chips that would allow it to comply with export controls, with the chipmaker keen to recapture a market that accounted for 26% of its turnover before the export rules, a number that has since dropped to around 17%.
The B20 would have to be far slower than the mainline version of Blackwell as U.S. export rules prevent companies from delivering chips with high chip-to-chip transfer speeds
Previous attempts by Nvidia to offer slower versions of its A100 and H100 GPUs to China fell foul of the export controls.
Senior analyst at Forrester, Alvin Nguyen added: “The difficulties that are being reported about Huawei’s software in its attempts to replace NVIDIA are unsurprising: the software ecosystem that NVIDIA has developed has been around for a long time (CUDA introduced in 2006) and NVIDIA continues to leverage this to their advantage.
"Huawei using their customer service capabilities to work more closely with customers in using their chips will help them counter NVIDIA’s advantage with their software ecosystem long-term.
“It is unrealistic to expect another organisation, like Huawei, to displace NVIDIA quickly – it will take time and effort from both a hardware and software perspective.
"Export controls will continue to give opportunities for Huawei to compete since the export control prevent NVIDIA from selling the “best” version of its products in China.”
RELATED STORIES
Nvidia shares fall despite record gains
