There's No Huawei Chinese Chipmakers Can Fill Nvidia's Shoes... Anytime Soon

Analysis Chinese chipmakers face an uphill battle filling the void left by the Biden administration's latest round of export restrictions, announced this week.

With the sale of most US-made AI accelerators and GPUs — including those developed by Nvidia and Intel for the Chinese market — soon to be restricted in the Middle Kingdom, meeting then demand is increasingly going to fall on domestic suppliers, who are still technologically playing catch up, or what's available elsewhere for HPC and AI functions.

As some, including the industry watchers at TrendForce have mused, Huawei's Ascend family of training and inference chips could see accelerated development in response to the American boycott. And, at least on paper, the cards do initially look promising as a replacement for US chips.

Introduced in 2019, the Atlas 300T training card and the Ascend 910 boasts FP16 performance within spitting distance of Nvidia's venerable A100. The full height, dual-slot card features 32GB of high-bandwidth memory and 16GB of DDR4 memory, and integrated 100Gb/s networking to scale out the system.

According to Huawei, the "Pro" version of the card is capable of 280 teraFLOPS of FP16. Assuming those figures aren't based on sparse mathematics — they don't say — that puts it 32 teraFLOPS behind Nvidia's A100 in terms of compute performance.

It's worth noting that Nvidia's H100 boasts FP16 performance 3.5x higher and substantially faster interconnect speeds to boot. So, by modern standards, it's not a particularly fast card.

However, Chinese chipmakers face numerous hurdles in bringing new AI accelerators to market. Among the biggest being who is going to make them?

While Huawei's Ascend 910 may look compelling, you might have picked up on the fact that, after four years, Huawei hasn't announced a successor. The reason is quite simple: the chip was built by Taiwan Semiconductor Manufacturing Co. (TSMC) — a company Huawei can no longer do business with after it landed on the US entities list in 2019.

A next-gen Ascend accelerator would have to rely on domestic semiconductor manufacturing capabilities, like those afforded by Semiconductor Manufacturing International Co. (SMIC). But, as we've previously reported, Huawei has been working on a homegrown Electronic Design Automation platform to assist with domestic production of chips, so it's a possibility.

Far and Huawei not the biggest problem

This problem isn't unique to Huawei.

In addition to restricting US export of most modern AI chips to the Middle Kingdom, the US also landed [PDF] 13 Chinese firms, including Biren Technology and Moore Threads, on the Commerce Department's "Entity List." Biren and Moore Threads are two of the more promising GPU vendors to come out of China in recent years.

Our sibling site The Next Platform took a close look at Biren's homegrown GPU at Hot Chips last year. The 77 billion transistor BR100 accelerator, built on a 7nm TSMC process, claims 1,024 teraFLOPS of BF16 performance — more than 3x that of Nvidia's A100. However, as we've previously reported, the chipmaker was forced to nerf the interconnect bandwidth last year to comply with a round of US export restrictions.

Moore Threads has focused more on the consumer GPU market, but has touted the use of its MTT S3000 cards for deep learning workloads. While details are a bit thin on the card's performance, the biz claims it's kit is capable of about 15.2 teraFLOPS of FP32 performance.

• The problem with Jon Stewart is that Apple appears to have cancelled his show
• Biden has brought the ban hammer down on US export of AI chips to China
• Fears China could trash Broadcom's VMware nuptials as revenge for sanctions
• America extends China chip export bans and acts to cut off backdoor exports

But just like Huawei's Ascend accelerator, both Biren and Moore Threads rely on foreign fabs, like TSMC, to build their chips for them. With both companies on the Entities List, it's unlikely TSMC and others will continue to service either of them.

Moving forward, development of new accelerators by Biren, Moore Threads, or Huawei will likely require refactoring their designs for production using older process tech available from domestic fabs like SMIC.

As we understand it, SMIC has only recently gained the ability to mass produce chips based on a 14nm process node. The chip maker has recently demonstrated, with the launch of Huawei's Mate 60 Pro, the ability to produce chips on a 7nm process.

Having said that, whether the foundry operator can actually do so in volume is up for debate. What's more, Chinese access to DUV equipment believed to have been used in the production of the chip will be severely limited once sanctions go into full effect next year.

Even if SMIC can keep up with demand, producing AI chips on any process node is substantially harder than a mobile SoC. AI chips tend to be much larger — Nvidia's H100 for instance is pushing the limits of this — and that means fewer dies per wafer. Depending on SMIC's 7nm defect rates, a single wafer might yield only a handful of usable parts — if any at all.

So, even if Huawei or Biren can adapt their designs for production on SMIC's 7nm process, there's a good chance the yield rates will be terrible.

While there's no question Chinese semiconductor manufacturing capability will improve over the next few years, domestic chipmaker's ability to fill the void left by US trade restrictions in the near term is going to be limited, especially once existing stock of inventory is exhausted. ®