Comment by ZenoArrow 3 days ago
I'd see Chinese RAM manufacturers like CXMT filling the void left in the market for consumer-grade RAM modules, I appreciate they face challenges (like lack of access to cutting edge EUV machines), but the RAM just needs to be fast enough and affordable enough for the average user for these companies to make significant inroads into the market that Micron, Samsung and SK Hynix are abandoning to chase the AI server market.
They will grow exponentially and catch the western market unawares in 10-15 years with a sudden flood of cheap, effective chips. Just like everything else China makes. Electric vehicles for example.
Sure, if they've got production grade EUV, but right now they don't even have production grade DUV.
I'm also sure they can go as far as 5nm like SMIC if they really wanted to, since it's strategic for China, but the cost would only be justified if the current cycle lasts long enough.
I was corrected elsewhere when I thought RAM was more expensive 10 years ago. RAM was actually cheaper 10 years ago, when it was DDR3/DDR4 too. If any company can replicate the 10 year old SOTA, they can bring prices down.
Just before the ASML ban, China imported as many DUV machines as possible.
China has a luxury of being able to not really care about the cost when it comes to what they view as a strategic advantage.
The problem is that the chances of the bursting of AI bubble seem far more likely than this happening first which you say about 10-15 years.
plus, the ram manufacturer cycle moves and does this all the time.
Atrioc does a really good job explaining these cycles[0]
But the point is that AI demand peaked when the supply was at its lowest which is why we are caught up in this messed up timeline that we live in. And this has sort of happen in the past too and this industries notorious for it (again watch the video, definitely worth it imo)
But still it feels like we are in this atleast for a year or two hard. Micron is iirc like suggesting what hundreds of billions of $ in factory investment right now and saying that the fastest might open in early 2027
Some estimates 2028 idk, I do feel like the chances of AI bubble popping around this time are likely too.
But still for atleast 1-2 years, we either as consumers or as small vps providers (yes the people who create vps providers are same people like you and me) are absolutely f*ed and the question is around that imo.
[0] The AI Tax Has Started: https://www.youtube.com/watch?v=nipeaKC3dWs
The thing is, even if the market bursts, prices are already inflated. RAM manufacturers know that everything 5xed and they aren't likely going to rush out and drop the price levels to pre-expansion. Once the AI market bursts, you can expect slow and methodical decreases in price (if any).
And that will ultimately buy China a lot of time to shove their ram into the market cutting ram manufacturers out of most non-US markets.
I think the major memory manufacturers are simply banking on their ability to flood the market if worst comes to worse. That or I could see some standards trickery around DDR6 (or some new BS standard). It'd not shock me if they coordinated with AMD/Intel to keep the standard secret as long as possible simply give themselves a lead in production.
> die size is 40% bigger than Samsung
likely naive question: why it is a problem? I would be fine if RAM in my PC has 10 times larger physical size if it is overall cheaper.
I guess that larger may have more power draw, but given costs of RAM and electricity and power draw of RAM it sounds unlikely to be a problem.
At a high end it would run into real-estate prices - at some point using half of room for computer stops making economical sense, given costs of rent or buying flat space. But just doubling size of PC does not sound like a bad tradeoff if it would be say 20% cheaper. Or 50% cheaper.
Is it about not fitting existing motherboards?
Is there reason why they cannot just make memories physically larger? It is "only" 40%, not 40000%
They're already producing 10nm DRAM with their current nodes, and they're working on producing 3d DRAM which may make node size somewhat moot.
Not 10nm, they are producing with 18.5nm and 17nm now, which technically already is in breach of US restrictions, the US government can blacklist them if they feel like it.
3D DRAM is no magic, it will only give them maybe 2 generations' breathing room if they got the required etching equipment figured out. But others will be doing 3D DRAM with EUV by then.
Are you sure?
>CXMT has begun mass production of DRAM using a D1z (sub-16nm) process.
https://www.globalsmt.net/advanced-packaging/decoding-cxmt-d...
They call it "10nm class" later in the article.
It's hard to find much concrete info tbh.
If the US government bans the import of RAM, it guarantees the immediate collapse of the US.
> Their scale is simply too small to affect the market outside China, majority of their chips will be eaten up by HBM3 production with yet unknown yield rate.
They don't currently have the tech to compete on HBM3 production, but they can produce DDR5 memory, and they will undoubtedly be scaling up production on this.
> which is already obsolete, die size is 40% bigger than Samsung, not to mention lithography, the big 3 are using EUV while they are stuck with lobotomized DUV.
And? That's good enough. My daily driver desktop, which I use to do development, plus play a few games (FC5, Dirt Rally, etc) has 16GB DDR3.
For 90% of computers in use today, including laptops, that RAM you call obsolete is fine.
Users aren't going to complain that a document which takes 3s to open now takes 3.5s
What you need as far as RAM goes, to make the computer perform acceptably is capacity. Users get a bigger performance boost by going from 8GB RAM to 16GB Ram and from 16GB Ram to 32 GB RAM than from DDR3 to DDR4 or from DDR4 to DDR5.
> They can squeeze out a bit more performance if they are ready to go beyond their current node using only domestic equipment and be blacklisted by the US government.
Which suits the rest of the world just fine. More for the rest of us, and if the single-digit-percent portion of their market that the US represents wants to lock itself out, no skin off anyone else's nose.
>market that Micron, Samsung and SK Hynix are abandoning to chase the AI server market
These three have collectively committed what, approaching $50B towards construction of new facilities and fabs in response to the demand?
The memory industry has traditionally projected demand several years out and proactively scheduled construction and manufacturing to be able to meet the projected demand. The last time they did that, in the crypto boom, the boom quickly turned into a bust and the memory makers got burned with a bad case of oversupply for years. With that context, can you blame them for wanting to go a bit more slowly with this boom?
Sure, the new fabs won't be up and at volume production until late 2027 / early 2028, but committing tens of billions of dollars to new production facilities, including to facilities dedicated to DRAM rather than NAND or HBM, is hardly 'abandoning'. They're pivoting to higher profit margin segments - rational behavior for a for-profit corporation - but thanks to the invisible hand of the (not quite as free as it should be) market, this is, partially, a self-solving issue, as DRAM margins soar while HBM margins compress, and we're already seeing industry response to that dynamic, too: https://www.guru3d.com/story/samsung-reallocates-of-hbm3-cap...
> Sure, the new fabs won't be up and at volume production until late 2027 / early 2028, but committing tens of billions of dollars to new production facilities, including to facilities dedicated to DRAM rather than NAND or HBM, is hardly 'abandoning'.
Look at what happened to Crucial. Why would Micron axe it's whole consumer RAM division if it was just experiencing a temporary drop in DRAM supplies until new fabs were brought online? Samsung and SK Hynix may have changes in priorities in the coming years, and in the case of Samsung I'm sure they'll still make sure to supply sufficient DRAM chips for the devices it manufactures (phones, TVs, etc...) but Micron has made it's current intentions fairly clear. They'll probably work with OEMs, but they're unlikely to return to selling to the general public any time soon.
What makes you think it won't start selling it as Micron brand or even Crucial brand once the bubble pops?
What makes you think consumers would want to trust them again after they were abandoned previously? Consumers would rather vote with their wallets for companies that are going to continue giving them good deals. The only reason Micron has a chance to pivot back to consumer RAM is because there's not much competition, but that could change if the Chinese RAM manufacturers can make inroads into western markets and continue their rapid technological growth.
This is the classic commodities cycle, and it happens everywhere in an economy where aggregate supply is inelastic in the short term but aggregate demand can fluctuate quickly. The reason it’s coming for DRAM first is that memory is the closest part to a pure interchangeable commodity and that recent process nodes have had almost zero improvement to memory density for years now, despite logic density continuing to increase exponentially. That and these companies have been known to fix prices in the past, but in this case the evidence suggests it’s a large aggregate demand shock.
China also needs RAM for AIs, especially since they have plenty of electrical power and building speed to pump out data-centers.
Turns out their wind "opercapacity" maybe isn't. Maybe they are trading chip efficiency for raw power.
Something I’ve been sort of wondering about—LLM training seems like it ought to be the most dispatchable possible workload (easy to pause the thing when you don’t have enough wind power, say). But, when I’ve brought this up before people have pointed out that, basically, top-tier GPU time is just so valuable that they always want to be training full speed ahead.
But, hypothetically if they had a ton of previous gen GPUs (so, less efficient) and a ton of intermittent energy (from solar or wind) maybe it could be a good tradeoff to run them intermittently?
Ultimately a workload that can profitably consumer “free” watts (and therefore flops) from renewable overprovisioning would be good for society I guess.
First: Almost anything can be profitable if you have free inputs.
Second: Even solar and wind are not really "free" as the capital costs still depreciate over the lifetime of the plant. You might be getting the power for near-zero or even negative cost for a short while, but the power cost advantage will very quickly be competed away since it's so easy to spend a lot of energy. Even remelting recycled metals would need much less capital investment than even a previous-gen datacentre.
That leaves the GPUs. Even previous gen GPUs will still cost money if you want to buy them at scale, and those too depreciate over time even if you don't use them. So to get the maximum value out of them, you'd want to run them as much as possible, but that contradicts the business idea of utilizing low cost energy from intermittent sources.
Long story short: in might work in very specific circumstances if you can make the numbers work. But the odds are heavily stacked against you because typically energy costs are relatively minor compared to capital costs, especially if you intend to run only a small fraction of the time when electricity is cheap. Do your own math for your own situation of course. If you live in Iceland things might be completely different.
I think as such intermittent power comes on the grid in the coming decades, people will find creative uses for it.
> top-tier GPU time is just so valuable that they always want to be training full speed ahead.
I don't think this makes much sense because the "waste" of hardware infrastructure by going from 99.999% duty cycle to 99% is still only ~1%. It's linear in the fraction of forgone capacity, while the fraction of power costs you save from simply shaving off the costliest peaks and shifting that demand to the lows is superlinear.
I truly wish Chinese Ram manufacturers luck to fulfill this market. Seriously, the amount of ram and its downstream effects can be hardly understated imo. it really just starts impacting everything.
That's probably what is going to happen, it's a strategic opportunity for the Chinese government here, there's a big market demand that can fuel their domestic production capabilities that nobody wants to take.
It would be a strategic opportunity for Intel, if they weren't run by imbeciles. DDR4 doesn't require the latest and greatest nodes. It's boring old technology. Even DDR5 is pretty boring. Intel could clean up fabbing DRAM (like they used to). But alas no. They're part of the semiconductor cartel and uninterested in the supply of DRAM increasing. Prices would drop and the fabs would only make stupid margins instead of disgusting margins.
Intel would be in the very same bind as every other DRAM producer who's trying to expand production today, only far worse because they have practically no experience fabbing DRAM compared to logic. (You can fab eDRAM on logic processes, but you'd only do that out of sheer desperation since the cost per bit is much higher.)
What they could do very easily in this market is bring back frickin Optane and hook it up to a modern PCIe bus with modern PCIe performance.
A recent development is High Bandwidth Flash (HBF) which works the same as HBM but uses less power and has far higher density.
If Intel made an equivalent high bandwidth Optane product, they’d immediately corner the GPU memory market.
It’s fine if it’s just filling Chinese manufacturing. Low-cost VPS hosts are going to be using brands like Supermicro anyways. It still gets exported.
Except for RAM from YMTC, which the USA gave a near-death sentence to by placing it on the Dept. of Commerce “Entity List” so no USA-associated business can do business with YMTC now.
This is true.
We use ASRock Rack servers, mainly because the only option for our industry are OEMs like Supermicro and ASRock. Dell and HPE are non-starters, except for our "storage" offering.
Back in 2019, HPE was a good midrange option. Then came ASRock Rack who obliterated HPE with the X470D4U, relegating HPE to high-end enterprise servers. But also made Ryzen-based VPS hosts including yours truly, BuyVM, et al.
I am very hopeful of CXMT. But then then it could take a while for them to ramp up production. Maybe by then, the AI bubble would've burst.
One problem with US sanctions is it could hurt US companies too, like in the case of cutting-edge EUV and CXMT. This is when China is actually a hero and not a villain.
We can certainly do with less plastic junk and fast fashion. But on the high end it hard to argue that cheaper Chinese products are ever a bad thing.
If corporations in western (aligned) countries stopped feeding sovereign wealth funds and private equity with profits and actually invested something maybe they could compete with China more closely, even with whatever shenanigans the CPC get up to with state support.
Their scale is simply too small to affect the market outside China, majority of their chips will be eaten up by HBM3 production with yet unknown yield rate.
They are forbidden to buy foreign equipment beyond their current process node, which is already obsolete, die size is 40% bigger than Samsung, not to mention lithography, the big 3 are using EUV while they are stuck with lobotomized DUV.
They can start making some decent money now, but vastly expanding capacity as is means enormous losses if the cycle went downward a few years later, that's how all previous makers went bankrupt.
They can squeeze out a bit more performance if they are ready to go beyond their current node using only domestic equipment and be blacklisted by the US government.
But the cap is there, unless they can make a working EUV machine in 5 years, they are doomed to be a minor player, if the current cycle even lasts that long.