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.
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.
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."
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.
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.
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.
Sure, designing modern integrated circuits isn't easy. However it still is way easier than what you need to manufacture them. Design of digital integrated circuits is commonly more understood with information being readily available.
In theory you could gain the knowledge to design an early 1990s CPU at the logic gate level by reading some books and doing a bit of research, on the other hand actually manufacturing such IC would take considerably more effort.
1. Why will they master it? Because they dedicate their industrial strategy and hence resources to it like they did in the other technological domains and flood the market?
2. Is the only way out a strict decoupling from the Chinese market in these domains? Or would it be a strategy that involves protecting domestic industries with other levers?
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.
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.
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.
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.
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.
> This feels comparable to the intel's battlemage offering of GPUs.
I'd very much like whatever card can run LLMs with Ollama or vLLM without bankrupting me and hopefully with somewhat low power usage.
Nvidia L4 cards seem to fit the bill when it comes to the power usage and getting things done, but the costs are way out there, not functionally different from H100s (I can afford neither).
So I'd very much welcome the Intel B60 Pro cards or honestly anything I could actually buy online. Until then, I'm stuck throwing money at OpenRouter and other API providers every month.
> 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.
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.
Many are so focused on the narrative that the media runs with so that they seem all but blind to see the more logical macro-view of the situation which is just sitting there in plain sight..
1. If you look at a world map you will quickly notice that the island of Taiwan is adjacent to the asian subcontinent and mainland China.
If you had one of those physical globes with a world map on them, you would not even see the United states of America when looking at Taiwan.
This fact alone should make you realize that the US has no business whatsoever interfering or warmongering about anything going on over there.
2. The reason for all this is to, by any means, try prevent any increase in "power" for other countries in an attempt to prevent their time in the sun on the global economic field from fading away.
3. The reason why this will fail is because the strategy of trying to force others to be your friends and to blackmail/strongarm them into submissive allies does not work in the long run.
Anyone who has attended the social battleground of kindergarten should have seen this to be a human fact.
4. My tip to all world leaders is as follows. Try not act like hyperagressive mentally challenged warmongering pssies and everything will be fine.
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.
Readit News
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."
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.
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 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.
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.
In theory you could gain the knowledge to design an early 1990s CPU at the logic gate level by reading some books and doing a bit of research, on the other hand actually manufacturing such IC would take considerably more effort.
Since you are from this domain:
1. Why will they master it? Because they dedicate their industrial strategy and hence resources to it like they did in the other technological domains and flood the market?
2. Is the only way out a strict decoupling from the Chinese market in these domains? Or would it be a strategy that involves protecting domestic industries with other levers?
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'd very much like whatever card can run LLMs with Ollama or vLLM without bankrupting me and hopefully with somewhat low power usage.
Nvidia L4 cards seem to fit the bill when it comes to the power usage and getting things done, but the costs are way out there, not functionally different from H100s (I can afford neither).
So I'd very much welcome the Intel B60 Pro cards or honestly anything I could actually buy online. Until then, I'm stuck throwing money at OpenRouter and other API providers every month.
Memory bandwidth for the M4 Max is 410 GB/s with the 32-core GPU or 546 GB/s with the 40-core GPU. It’s 819 GB/s in the M3 Ultra config.
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.
Support for ECC
@150Watts
1. If you look at a world map you will quickly notice that the island of Taiwan is adjacent to the asian subcontinent and mainland China. If you had one of those physical globes with a world map on them, you would not even see the United states of America when looking at Taiwan. This fact alone should make you realize that the US has no business whatsoever interfering or warmongering about anything going on over there.
2. The reason for all this is to, by any means, try prevent any increase in "power" for other countries in an attempt to prevent their time in the sun on the global economic field from fading away.
3. The reason why this will fail is because the strategy of trying to force others to be your friends and to blackmail/strongarm them into submissive allies does not work in the long run. Anyone who has attended the social battleground of kindergarten should have seen this to be a human fact.
4. My tip to all world leaders is as follows. Try not act like hyperagressive mentally challenged warmongering pssies and everything will be fine.
/Michael Ah. Sweden
I guess you could try to make a GPU with just programmable compute but Intel attempted that and it didn't go very well.
Deleted Comment