Naive question: Are the current (from what I have heard not very effective) export restrictions of HPC GPUs to china truly productive in the long run if the goal is to retain an edge? As in, to me it seems that it just fuels an expansion of domestic capabilities and in the car and solar sector my impression is that china had already proven that it can absolutely perform on par or even better in many different metrics compared to western countries, given time and pressure. So while these chips are not on par with current or even last gen GPUs, I would not be surprised if china would catch up and even have a much higher incentive to do so, now that other countries try to control their access to key technologies.
I am not saying whether retaining an edge is good or bad or that I have a different answer if one thought it was good. Just curious what you guys think.
sho_hn · 6h ago
I would not be surprised if most of us are running Chinese silicon a decade or two from now, unless China invades Taiwan, and I also think recent events have certainly spurned CCP tech strategy and accelerated this timeline.
There's a few hurdles for China to overcome first, most notably catching up on high-end manufacturing processes, but it's naive to assume that won't happen eventually.
For consumer and prosumer gear that they can get it done is already obvious, cf. people generally having no problem with buying DJI, BambuLabs or Anker.
bbarnett · 6h ago
China will 100% invade Taiwan. This is why both parties in the US are spending so much to get domestic chip production running.
I would be astonished if the backroom deal wasn't "If you take Taiwan now, we'll have to stop you, if you wait until we're self sufficient, we won't interfere."
This is the strategy for decades but it’s failed to produce results
lossolo · 4h ago
The demographic situation in Taiwan is collapsing. Taiwan is basically fully dependent on trade with China and so on. The Chinese are masters of long term strategy and patience and would rather use deception than a sword if possible.
lossolo · 4h ago
> China will 100% invade Taiwan.
Stop looking at China through a Western lens. No one knows what China will do, so this statement is false. Considering their history and culture, they will first use all other tactics to take over the island. They said they want to do it by 2049, and they could succeed without firing a single bullet, as the commenter below noted.
> I would be astonished if the backroom deal wasn't "If you take Taiwan now, we'll have to stop you, if you wait until we're self sufficient, we won't interfere."
None of the cutting edge nodes are, or will be, produced in the US by TSMC. They are all produced in Asia, and the fabs in US will be X years behind.
xethos · 2h ago
> None of the cutting edge nodes are, or will be, produced in the US by TSMC. They are all produced in Asia, and the fabs in US will be X years behind.
I don't doubt you, and I think you'll be right for some time.
I also think it's foolish to count out Intel. They're down, but so was AMD. More pointedly, this is not Intel's first time playing "This works or we go under"
chrsw · 6h ago
China is racing full speed ahead to win in all these tech domains regardless of export controls.
They will surpass us on chips just like they surpassed us on EVs. The leading edge of chip design is very complex so it will just take more time than EVs. But it is inevitable.
Even if China could get their hands on all the NVIDIA GPUs they wanted they would still try to make their own as fast as possible.
tangotaylor · 6h ago
I think it will be effective. This stuff is hard. There used to be many competitors capable of the best process technology: TI, GlobalFoundries, Intel, IBM, Samsung, TSMC.
Canon, Nikon, ASML all used to have competitive lithography machines.
Now it’s just TSMC and Samsung at the edge, and only ASML supplies the latest lithography machines.
China will probably catch up quickly but the pace will be nonlinear and illusory. They will hit diminishing returns just like everyone else has.
They’ve probably stolen every bit of semiconductor IP they can through economic coercion or espionage.
All they can do now is out-innovate everyone else and that will take a long time. But who knows, their pace of advancement since Mao died has been impressive.
trm42 · 11m ago
One interesting detail is that the Chinese have been improving their photography lens production and quality in rapid pace and cheap price.
The legendary Zeiss is producing the lithography lenses for ASML, so it looks like China is pouring lots of effort to photography lenses to bootstrap their lithography lens capabilities.
I don’t know about the other parts needed for chip fabbing but I kinda expect then to encourage and subsidize other technological fields related to it as well.
ac29 · 5h ago
> Now it’s just TSMC and Samsung at the edge
Intel 3 has been shipping since last year and is only very slightly behind TSMC N3.
TSMC is almost certainly doing far more volume on their leading node though.
sciencesama · 5h ago
Smic is led by the person who spear headed tsmc !
trasirinc · 6h ago
That's assuming they can keep pumping massive capital into every industry that it seeks to circumvent from bans and sanctions. But it appears they have very short runway these days. Just months after the initial tariffs/sanctions from US, Chinese government is enacting multiple tax raising schemes in September to try to stay alive. The first is the mandating that workers and employees cannot opt out of social security contributions. which is around 1500 yuan ($200) per month for one worker. for an average worker that makes 4000 ($600) yuan, it makes no sense. So many companies are deciding to layoff or close up in September. And workers are going back to countryside. The second is the landlord tax that is starting on September 15th. This is due to people not buying real estates anymore and renting instead.
stogot · 4h ago
Are you a local? What city?
HAL3000 · 4h ago
Langley
bb88 · 4h ago
We can look at history.
The US has export restrictions on certain computing devices to certain regimes which included the Sony Playstation 2, a gaming console from the double noughts [0]. Apparently the military thought it could be used to create nasty weapons. Two decades later and nobody cares whether a PS2 is shipped to Iran. We still track FPGAs I guess, though I haven't checked what's on the ITAR/EAR list in a while.
Embargoes typically work until the embargoees(?) develop the technology to build or acquire what they need. If AI is only a strategic advantage because of hardware alone, then yes. But Deepseek kinda maybe killed that idea. China has never been the first mover. They optimize. But it looks like today, AI embargoes to China will get the US months at most.
This feels comparable to the intel's battlemage offering of GPUs. Competitive when it comes to price, but simply not at the level of nvidia's offerings nor are they usable for AI.
I highly doubt these will overtake the modded 48gb 4090 usage in China, but still it's a clear indicator that the chip embargo has lit a fire under Chinese industry to accelerate their own offerings in the semiconductor sector. It'll be interesting to see how these evolve in the coming months/years, as I suspect their iteration cycles will be shorter than US counterparts.
jauntywundrkind · 5h ago
Huawei doing what the West won't, smashing the product segmentation niches. 96GB is what counts here. Being able to fit a good model is more important than speed, for many folks.
Also worth trying to better compare on efficiency. I don't know how this shakes up, but TOps/W is another figure of merit that also matters a lot, maybe more than absolute TOps count. I don't know how good 1.86TOps/W measures up here, but knowing that this is 150W is quite time where-as a 4090 or whatever is way way more reasonable than how most cards get built.
com2kid · 5h ago
You can do a lot with a lot of VRAM and a reasonable amount of TOPS.
A 24GB affordable GPU can easily power an entire house worth of AI work, from real time voice chat, image generation, simple tool calls and task running, reminders, alerts, smart home integrations, etc.
IMHO a large set of potential use cases is being held back by Nvidia's high prices.
sho_hn · 6h ago
This is a first-gen effort from Huawei, no? Plenty of resources and time to iterate.
wakawaka28 · 6h ago
Isn't it also an early effort for Intel? They are not known for their GPUs. Also, I'm sure Huawei has some chip tech, and GPUs are mostly just a ton of little processors arranged in a grid.
GodelNumbering · 7h ago
LPDDR4X 96GB, total bandwidth 408GB/s
jsheard · 7h ago
For comparison, Nvidia's 96GB card uses GDDR7 for 1.8TB/sec bandwidth. This Huawei card is more in the league of an M4 Max bandwidth-wise.
sciencesama · 7h ago
But the price !
nine_k · 6h ago
I wonder if Huawei are making any money off this, or are they trying to get a foot in the door first.
moralestapia · 6h ago
Which is?
userbinator · 6h ago
If they release the full hardware documentation they will already be ahead of the incumbent in openness, just like Intel and AMD did; and IMHO that will also help gain a following.
wakawaka28 · 6h ago
If openness becomes the killer feature, their competitors will also release docs or whatever else.
commandersaki · 7h ago
I'm not really familiar with GPUs that are not Intel, AMD, and Nvidia, but would these work fine for gaming (realising it is not the primary purpose).
smallmancontrov · 7h ago
Gaming has an enormous API surface compared to AI, so almost certainly not.
jsheard · 7h ago
Gaming also wants a bunch of fixed-function graphics units that a dedicated accelerator like this has no reason to include.
I guess you could try to make a GPU with just programmable compute but Intel attempted that and it didn't go very well.
ZiiS · 7h ago
They would need very complex drivers writing.
abracadaniel · 7h ago
At this point it seems like most of what goes into the driver must be per-game tweaks to fix mistakes or optimize unoptimized code.
chickenzzzzu · 7h ago
Thats from the opengl days, that has mostly been shifted to vulkan and dx12 now. still, you end up having the hw vendor employees do the work in the game company's codebase
washadjeffmad · 7h ago
Seems like not a bad add-in card for OpenCV/2.
sciencesama · 5h ago
There was an nvidia hack that was related to china, they got all the technology they need now !
OsrsNeedsf2P · 7h ago
Certainly big for China, but if it can't run CUDA then it's not going to help them catch up in the AI race
LeoPanthera · 7h ago
Macs can't run CUDA but everyone is buying those for AI anyway.
jszymborski · 7h ago
GP was likely referring to training, not inference.
wolfgangK · 7h ago
Only those who don't care/know about prompt processing speed are buying Macs for LLM inference.
com2kid · 5h ago
Even 40 tokens per second is plenty enough for real time usage. The average person reads at ~4 words per second, 40 tokens per second is going to be 15-20 words per second.
Even useful models like gemma3 27b are hitting 22 t/s on 4bit quants.
You aren't going to be reformatting gigabytes of PDFs or anything, but for a lot of common use cases, those speeds are fine.
esseph · 6h ago
Don't know and don't care are definitely things that I could be, but it also makes sense if they want to keep lookups private.
sliken · 4h ago
Dunno, training maybe, for inference pytorch and llama seem more important.
chermi · 6h ago
I thought they were already working around cuda?
aswanson · 7h ago
Maybe they'll make a better abstraction layer than cuda.
chickenzzzzu · 7h ago
Vulkan already achieves 95% of cuda, with the remaining 5% being scheduling.
atlgator · 6h ago
How many do I need to run the full DeepSeek 671B?
robotnikman · 5h ago
Using a rough estimate where 1B parameters requires 1GB of VRAM, you would need at least 7
nextworddev · 6h ago
Just a matter of time till they steal the recipe
jiggawatts · 7h ago
So… five to ten times worse TOPS/W compared to NVIDIA, if I’m reading this correctly.
It wouldn’t have a market if it wasn’t for the tensions between the USA and China, unless it’s super cheap.
diggan · 6h ago
> It wouldn’t have a market if it wasn’t for the tensions between the USA and China, unless it’s super cheap.
It is super cheap, compared to what's available if you want 96GB for ML inference in a single card (ignoring the other aspects one might care about). I'm seeing it on Alibaba for 1200-1500 EUR which is like 7-8 times cheaper than I can buy a RTX Pro 6000 for locally.
enlyth · 7h ago
It has shitty LPDDR4X memory
cuuupid · 7h ago
I predict the tech reaction will be underwhelmed/disappointed and the wall street reaction will be panic selling NVDA
kittikitti · 7h ago
This is really great. Things like CUDA support are not a requirement for my purposes. The DDR4 speeds seem like a bottleneck but the VRAM in GPU's are most often also a bottleneck. I look forward to more technical reports on their AI accelerator.
rvz · 7h ago
No need to panic, given that it is not compatible and cannot run CUDA....yet.
But don't ignore this at all. This industry can change in 1 or 2 years very quickly.
rowanG077 · 7h ago
Why panic? It's great that finally there is AI hardware coming out of china. This can only be good for everyone.
beeflet · 7h ago
except for humans
artninja1988 · 6h ago
Why? I'm a human and I like competitive markets and hate monopolies
beeflet · 6h ago
What's your opinion on being tortured by a machine for eternity?
Ifkaluva · 6h ago
Ok doomer
beeflet · 3h ago
If it's not OK then it's not the end
hereme888 · 7h ago
Let's see how it performs in the real world. The website makes some outlandish claims, like 1.86 TOPS/W leads the industry... NVIDIA's comparable RTX 6000 Pro gives out 6.67 TOPS/W.
Qualcomm's Cloud AI 100 Ultra has like 100 TOPS/W.
koolala · 7h ago
Why is 1.86 TOPS/W outlandish? Why does 1.86 TOPS/W lead the industry?
sho_hn · 7h ago
I think they are saying the website claims that 1.86 TOPS/W is industry-leading while it is not.
Insanity · 7h ago
I guess the outlandish claim is that it’s industry leading while it’s not, as other GPUs outperform it.
hereme888 · 7h ago
Because, as I showed, 1.86 is nowhere near industry-leading.
trasirinc · 7h ago
It's considerably slower than RTX 4090, doesn't have CUDA, and you are entirely beholden to whatever sketchy drivers they have. And power consumption is generally also dog shit.
There are YouTubers who have bought other cards to test them out, and drivers are generally the big problem.
Chinese hardware manufacturers usually only target and test on the hardware/software configs available in China. They mostly use the same stuff, but with weird quirks due to Chinese ownership and modification of a lot of stuff that enters their country.
You can't seriously expect a new GPU manufacturer to create a perfectly useable ecosystem on day one.
The drivers will surely get better over time and support for integrating the compute stack that they use will come if the incentive is good enough.
I really hope this doesn't turn out like HIP in AMD Radeon cards. That is absolute dog shit and has been dog shit for ages. It's really sad that an AMD card from 2017 is useless for compute while an equivalent Nvidia card from the same era is just now getting dropped by the latest CUDA versions.
echelon · 6h ago
Maybe the Chinese can break the CUDA monopoly?
They're the ones writing most of the open source AI code.
trasirinc · 6h ago
Because it's not just "CUDA" by itself, it's the whole ecosystem.
Most AI developers don't actually use CUDA directly, they use programming libraries like PyTorch that use CUDA under the hood to communicate with the GPU to run tasks in parallel.
CUDA is pretty much the standard and is supported anywhere it's relevant.
Just creating an alternative is pretty meaningless if it isn't actually supported anywhere.
Adding support isn't easy, and there's also stability issues, bugs, etc. People want something that works and is reliable (= CUDA, since it's battle-tested).
That's the same flawed argument that people have used to expect Huawei to replace Android/EUV.
xadhominemx · 6h ago
Extremely large and fragmented community of AI researchers and developers in China. CUDA competitors in China face the same hurdles as those in the West.
I am not saying whether retaining an edge is good or bad or that I have a different answer if one thought it was good. Just curious what you guys think.
There's a few hurdles for China to overcome first, most notably catching up on high-end manufacturing processes, but it's naive to assume that won't happen eventually.
For consumer and prosumer gear that they can get it done is already obvious, cf. people generally having no problem with buying DJI, BambuLabs or Anker.
I would be astonished if the backroom deal wasn't "If you take Taiwan now, we'll have to stop you, if you wait until we're self sufficient, we won't interfere."
Stop looking at China through a Western lens. No one knows what China will do, so this statement is false. Considering their history and culture, they will first use all other tactics to take over the island. They said they want to do it by 2049, and they could succeed without firing a single bullet, as the commenter below noted.
> I would be astonished if the backroom deal wasn't "If you take Taiwan now, we'll have to stop you, if you wait until we're self sufficient, we won't interfere."
None of the cutting edge nodes are, or will be, produced in the US by TSMC. They are all produced in Asia, and the fabs in US will be X years behind.
I don't doubt you, and I think you'll be right for some time.
I also think it's foolish to count out Intel. They're down, but so was AMD. More pointedly, this is not Intel's first time playing "This works or we go under"
They will surpass us on chips just like they surpassed us on EVs. The leading edge of chip design is very complex so it will just take more time than EVs. But it is inevitable.
Even if China could get their hands on all the NVIDIA GPUs they wanted they would still try to make their own as fast as possible.
Canon, Nikon, ASML all used to have competitive lithography machines.
Now it’s just TSMC and Samsung at the edge, and only ASML supplies the latest lithography machines.
China will probably catch up quickly but the pace will be nonlinear and illusory. They will hit diminishing returns just like everyone else has.
They’ve probably stolen every bit of semiconductor IP they can through economic coercion or espionage.
All they can do now is out-innovate everyone else and that will take a long time. But who knows, their pace of advancement since Mao died has been impressive.
The legendary Zeiss is producing the lithography lenses for ASML, so it looks like China is pouring lots of effort to photography lenses to bootstrap their lithography lens capabilities.
I don’t know about the other parts needed for chip fabbing but I kinda expect then to encourage and subsidize other technological fields related to it as well.
Intel 3 has been shipping since last year and is only very slightly behind TSMC N3.
TSMC is almost certainly doing far more volume on their leading node though.
The US has export restrictions on certain computing devices to certain regimes which included the Sony Playstation 2, a gaming console from the double noughts [0]. Apparently the military thought it could be used to create nasty weapons. Two decades later and nobody cares whether a PS2 is shipped to Iran. We still track FPGAs I guess, though I haven't checked what's on the ITAR/EAR list in a while.
Embargoes typically work until the embargoees(?) develop the technology to build or acquire what they need. If AI is only a strategic advantage because of hardware alone, then yes. But Deepseek kinda maybe killed that idea. China has never been the first mover. They optimize. But it looks like today, AI embargoes to China will get the US months at most.
[0] https://www.pcmag.com/news/20-years-later-how-concerns-about...
I highly doubt these will overtake the modded 48gb 4090 usage in China, but still it's a clear indicator that the chip embargo has lit a fire under Chinese industry to accelerate their own offerings in the semiconductor sector. It'll be interesting to see how these evolve in the coming months/years, as I suspect their iteration cycles will be shorter than US counterparts.
Also worth trying to better compare on efficiency. I don't know how this shakes up, but TOps/W is another figure of merit that also matters a lot, maybe more than absolute TOps count. I don't know how good 1.86TOps/W measures up here, but knowing that this is 150W is quite time where-as a 4090 or whatever is way way more reasonable than how most cards get built.
A 24GB affordable GPU can easily power an entire house worth of AI work, from real time voice chat, image generation, simple tool calls and task running, reminders, alerts, smart home integrations, etc.
IMHO a large set of potential use cases is being held back by Nvidia's high prices.
I guess you could try to make a GPU with just programmable compute but Intel attempted that and it didn't go very well.
Even useful models like gemma3 27b are hitting 22 t/s on 4bit quants.
You aren't going to be reformatting gigabytes of PDFs or anything, but for a lot of common use cases, those speeds are fine.
It wouldn’t have a market if it wasn’t for the tensions between the USA and China, unless it’s super cheap.
It is super cheap, compared to what's available if you want 96GB for ML inference in a single card (ignoring the other aspects one might care about). I'm seeing it on Alibaba for 1200-1500 EUR which is like 7-8 times cheaper than I can buy a RTX Pro 6000 for locally.
But don't ignore this at all. This industry can change in 1 or 2 years very quickly.
There are YouTubers who have bought other cards to test them out, and drivers are generally the big problem.
Chinese hardware manufacturers usually only target and test on the hardware/software configs available in China. They mostly use the same stuff, but with weird quirks due to Chinese ownership and modification of a lot of stuff that enters their country.
The drivers will surely get better over time and support for integrating the compute stack that they use will come if the incentive is good enough.
I really hope this doesn't turn out like HIP in AMD Radeon cards. That is absolute dog shit and has been dog shit for ages. It's really sad that an AMD card from 2017 is useless for compute while an equivalent Nvidia card from the same era is just now getting dropped by the latest CUDA versions.
They're the ones writing most of the open source AI code.
Most AI developers don't actually use CUDA directly, they use programming libraries like PyTorch that use CUDA under the hood to communicate with the GPU to run tasks in parallel.
CUDA is pretty much the standard and is supported anywhere it's relevant.
Just creating an alternative is pretty meaningless if it isn't actually supported anywhere.
Adding support isn't easy, and there's also stability issues, bugs, etc. People want something that works and is reliable (= CUDA, since it's battle-tested).
That's the same flawed argument that people have used to expect Huawei to replace Android/EUV.