There's no Huawei Chinese chipmakers can fill Nvidia's shoes... anytime soon
- Reference: 1697972411
- News link: https://www.theregister.co.uk/2023/10/22/huawei_nvidia_replacement/
- Source link:
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 [1]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 [2]mused , Huawei's Ascend family of training and inference chips could see accelerated development in response to the American boycott. And, at least on [3]paper , the cards do initially look promising as a replacement for US chips.
[4]
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.
[5]
[6]
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.
[7]
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 [8]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 [9]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
[10]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.[11]
Our sibling site The Next Platform took a [12]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 [13]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 [14]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.
[15]The problem with Jon Stewart is that Apple appears to have cancelled his show
[16]Biden has brought the ban hammer down on US export of AI chips to China
[17]Fears China could trash Broadcom's VMware nuptials as revenge for sanctions
[18]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 [19]produce chips on a 7nm process.
Having said that, whether the foundry operator can actually do so in volume is up for [20]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. ®
Get our [21]Tech Resources
[1] https://www.theregister.com/2023/10/19/china_biden_ai/
[2] https://www.trendforce.com/presscenter/news/20231019-11891.html
[3] https://e.huawei.com/en/products/computing/ascend/atlas-300t-training-9000
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2ZTVHK6yRLyQUDReXtDItRQAAAEY&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[5] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZTVHK6yRLyQUDReXtDItRQAAAEY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZTVHK6yRLyQUDReXtDItRQAAAEY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44ZTVHK6yRLyQUDReXtDItRQAAAEY&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[8] https://www.theregister.com/2020/02/26/huawei_google_embargo_exemption_plea/
[9] https://www.theregister.com/2023/03/24/huawei_eda_14nm_chips/
[10] https://www.bis.doc.gov/index.php/documents/federal-register-notices-1/3361-88-fr-71991/file
[11] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_onprem/systems&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33ZTVHK6yRLyQUDReXtDItRQAAAEY&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[12] https://www.nextplatform.com/2022/08/25/china-launches-the-inevitable-indigenous-gpu/
[13] https://www.theregister.com/2022/11/08/chipmakers_alter_products_china_ban/
[14] https://www.mthreads.com/product/S3000
[15] https://www.theregister.com/2023/10/20/jon_stewart_apple/
[16] https://www.theregister.com/2023/10/19/china_biden_ai/
[17] https://www.theregister.com/2023/10/19/china_vmware_broadcom/
[18] https://www.theregister.com/2023/10/18/america_china_sanctions/
[19] https://www.theregister.com/2023/09/19/huaweis_advanced_chip_sparks_us/
[20] https://www.theregister.com/2023/09/19/huaweis_advanced_chip_sparks_us/
[21] https://whitepapers.theregister.com/
Taiwan occupation?
I'm wondering if this could backfire. Hurting China like this must make invasion of Taiwan - if only to get their hands on advanced foundry technologies - a more tempting proposition.
Peace in that region is fragile and carefully balanced. Pushing China into a corner could tip that and Western tech companies are dangerously exposed to any instability involving Taiwan.
Re: Taiwan occupation?
Invading Taiwan is very unlikely to result in a working advanced foundry.
There are no economic gains for China invading when the consequences are thought-through, it would only be done due to political reasons and the worry is as Xi gets older he wants to be remembered for something great and forced reunification might fit that bill. However, the progress (or lack of) that Russia has demonstrated in taking Ukraine might make the generals think carefully. Yes, I am sure China has the military might to "win" but at what cost?
Competition
Chinese are smart people. Hopefully they will develop their own, better, technology.
After all there is nothing better than a good old cold war technology race.
Giving China everything on a plate would make everyone involved complacent.
I hear you...
If, according to you, "a good old cold war technology race" is the best way for mankind to progress, then let's do more: cut pharma, cut biotech, cut nanotech, cut robotech, cut github, etc...
Surely, the "smart people" will develop much better alternatives than slow free countries in all these domains...
You know, like the USSR did, before the fall of the Berlin wall...
Re: I hear you...
You know, like the USSR did, before the fall of the Berlin wall...
Soviets are somewhat primitive people. They ruled through use of violence and their "advancements" were driven by saving face rather than progress and mostly by captive population of Ukrainians.
I think situation with China is vastly different.
You may not agree with their version of communism, but they are certainly not primitive.
Re: I hear you...
> Soviets are somewhat primitive people. They ruled through use of violence
As opposed to China one assumes (Taiping, Nian, Boxers, Xinhai, Great Leap Forward, Cultural revolution, Tiananmen, etc...).
Also look up " [1]Chinese execution vans " and the link to [2]organ harvesting practices .
[1] https://en.wikipedia.org/wiki/Execution_van
[2] https://en.wikipedia.org/wiki/Execution_van#Organ_harvesting
There is little doubt
Chinese vendors only have two options. De-Americanisation and self sufficiency.
No one questions either of the two is possible or even happening. The second one (just as with the EU) was declared years before Trump started messing with global supply chains. It was simply declared using different terms (self sufficiency).
The first point is a result of radical weaponisation of technology and most analysts agree that de-Americanisation is currently underway outside China to some degree. Back in 2019 it was estimated that re-design and production could take between 3 and 5 years. That makes the first visible signs likely to start appearing around now.
Huawei itself has already de-Americanised over 13,000 components, pumping millions into its Chinese suppliers and requiring them to increase quality in the process.
It has also invested across the chip design and
manufacturing tool chain.
Each wave of US sanctions has cut chunks out of the revenues of US companies. They (and the associations that represent them) are not happy.
The recent 'boom' in earnings from some of them is simply the fruit of 'non-demand' items. Basically stockpiling efforts from Chinese companies.
Revenues across the entire US semiconductor industry will be impacted severely. Those revenues are required for future R&D. They have been impacted since 2019.
Now, it is Chinese companies that are seeing massive rises in investment and revenues through real demand (as opposed to non-demand orders) combined with government subsidies.
Huawei's Ascend line was more than AI inference from the outset.
It was a full stack solution across both hardware and software. Ascend processors scale down to earbuds and IoT. Paired with MindSpore and CANN it was always going to have legs and HiSilicon was going to guide it along the roadmap.
It is worth noting that, ever since the sanctions began, HiSilicon embarked on a recruitment program for top talent (paying industry leading salaries to attract talent from all over the globe).
It has not changed tack. The last hurdle of fabrication remains but China has now been forced to accelerate development in that area.
It would be foolish to speak of time frames with any degree of precision but that is what some did. In 2019 many industry watchers claimed it could take decades for China to catch up.
That is a bit crazy when we realise that we could very probably be on the cusp of a move beyond silicon and absolutely no one knows what China is cooking.
With that in mind all we can do is look at the bigger picture.
In the smartphone sector, Huawei has claimed that 'sanctions' are the new 'normal'. They have said that they are moving back to a two flagship per year release cycle (plus folding and flip). What does that indicate?
Qualcomm has stated that it will see no further material revenue from Huawei. What does that indicate?
We know that de-Americanisation is well underway across the supply chain.
We know that Huawei's biggest ever software project (MetaERP) was completed in record time with yet more revenue impact for an American company (Oracle in this case).
We know that it is moving forward with self developed alternatives to wireless device interconnection (NearLink).
We know that both HarmonyOS and EulerOS are moving in tandem to lay the foundations for self reliance in the OS space. HarmonyOS NEXT has already been announced (totally Android free). There are rumours of a desktop flavour of HarmonyOS for 2024.
These are just tiny pieces of a bigger puzzle but Huawei does appear to be reservedly optimistic.
Canon just broadsided ASML with its imprint system.
Could a Chinese effort deliver an equally groundbreaking solution to the fabrication issue?
And where do we think RISC-V chiplet solutions will be just 5 years from now?
Pangu?
What are the chances of Huawei or Chinese interests actually leapfrogging current technologies in the near term?
With US interests haemorrhaging revenues in the near term through the loss of major Chinese sales volumes, how will they fund R&D going forward. After all, it isn't really sanctions that matter here, but convincing sovereign nations to impose export restrictions. Taiwan and South Korea have already managed to get permanent exemptions. ASML is far from happy with the current geopolitical situation.
My gut feeling is that the fab problem is closer to being resolved than many believe.
If these sanctions are such a good think for Huawei...
... then why are they fighting tooth and nail against them?
I'm confused now.
Re: If these sanctions are such a good think for Huawei...
No need to be confused.
Huawei has no issues doing business with US companies. In the short term at least it makes good sense.
Good business sense for everyone.
The irony is that on the business side, in a global supply chain, everybody wants to do business with everyone else.
The problem is politics and the shortsightedness of those in power.
They will go down in history as accelerating China's march to self sufficiency and imposing enormous self harm on its semiconductor industry.
Every single US Huawei supplier has applied for a licence to do business with the company.
They know the dire outlook that being denied business options with Huawei/China will bring.
Huawei stopped fighting against sanctions long ago.
Now it is simply working to overcome them.
I am not well up on how to train an AI, but does it have that much of an impact if your not using cutting edge tech is from companies like NVIDIA or Intel?
So even if the Chinese companies have to use home grown kit which is only half as fast as what the US companies can offer, would it not just take longer to train them up their AI but they will still get their in the end?
And can it not be scaled up by throwing more GPUs at the issue to compensate for each individual ones not being as fast?
Oh no!!!
How are these CCP facial recognition platforms going to work now?
These Megvii, SenseTime, CloudWalk, Yitu and others?
Yet another hurdle on the long march towards universal communist nirvana... Too bad.