RAMageddon? 2027 Memory Capacity Reportedly Sold Out (ign.com) 133
IGN reports:
[A] new report suggests that all three memory manufacturers — Samsung, SK Hynix, and Micron — have collectively sold through their 2027 memory manufacturing capacity to, you guessed it, AI companies.
According to a report from Digitimes and spotted by TweakTown, all DRAM and HBM capacity has been sold through for 2027, with no further supply planned. (The companies themselves have not yet confirmed this to be the case.) Of course, that only means that retail prices are going to go up even further in the near future. Selling through all the capacity for next year would be bad enough for consumers, but the problem is that most of the companies buying up this capacity have done so through long-term purchasing agreements, essentially pre-ordering memory that hasn't even been manufactured yet over a five-year term. Unless more capacity opens up somewhere in the next few years, it's hard to imagine that RAM will become more affordable and less scarce any time soon.
It's not just memory that's affected. According to that Digitimes report, NAND storage is growing in demand too. Luckily, NAND capacity hasn't been completely absorbed, because there are more companies supplying the stuff. But it's not exactly hard to see that SSDs have been growing steadily more expensive over the last six months or so.
According to a report from Digitimes and spotted by TweakTown, all DRAM and HBM capacity has been sold through for 2027, with no further supply planned. (The companies themselves have not yet confirmed this to be the case.) Of course, that only means that retail prices are going to go up even further in the near future. Selling through all the capacity for next year would be bad enough for consumers, but the problem is that most of the companies buying up this capacity have done so through long-term purchasing agreements, essentially pre-ordering memory that hasn't even been manufactured yet over a five-year term. Unless more capacity opens up somewhere in the next few years, it's hard to imagine that RAM will become more affordable and less scarce any time soon.
It's not just memory that's affected. According to that Digitimes report, NAND storage is growing in demand too. Luckily, NAND capacity hasn't been completely absorbed, because there are more companies supplying the stuff. But it's not exactly hard to see that SSDs have been growing steadily more expensive over the last six months or so.
No RAM, dwindling customers for AI (Score:5, Insightful)
Re: (Score:3)
Re: (Score:2)
Re: (Score:2)
It's not. The IGN article the summary quotes carelessly implies that, but it's not true.
The Internet tells me that memory manufacturing normally has a lead time of 12-18 months. Guess what's 16 months from now? 2028.
Re: No RAM, dwindling customers for AI (Score:2)
I'm willing to bet dwll, Lenovo, apple, and Asus all have some of that capacity booked too.
Maybe even some retailers.
Re:No RAM, dwindling customers for AI (Score:5, Informative)
All that is true... But maybe it's time people started exploring the 2nd hand market.
Buying a new computer isn't always necessary. A computer from 2020 will do just fine for most people. And DDR4 ram is considerably cheaper than DDR5.
Maybe this IS the year of the Linux desktop, cause people can't afford to use anything else (given Windows 10 is EOL, and everyone and their aunt is finding bugs everywhere, and Windows 11 needs TPM == new hardware). Or you could also just buy an Apple Neo, good value these days.
Re: (Score:2)
The market will be flooded with used hardware when the big AI companies start to go bust or sell off their older hardware that is being replaced. It will have had a hard life in a server rack, but stuff like RAM rarely fails due to age or use, and even consumables like SSDs are often for consumers if they are enterprise grade with 5% life left. I've been using some old Intel server SSDs with 5-10% life left for years, for bulk temporary storage.
Re: (Score:3)
I've heard the RAM used in AI rack servers is packaged in a format that isn't useful for generalist PC use. I don't know if that means standard chips surface mounted onto the main board that could be reflowed onto a standard DIMM or if the pin-out of the package is totally incompatible. I hope you are right, but I don't know if it will be the windfall for enthusiast PC builders we're hoping for.
I wouldn't mind having a whole ass compute unit or two to run my own local open weight model though if they ever g
Re: (Score:3)
AI uses a lot of HBM (High Bandwidth Memory) which is optimized for throughput, not latency. You are right, it's generally not something you would use for main RAM, it's for GPUs.
Re: No RAM, dwindling customers for AI (Score:2)
Well also HBM sits on top of a compute die work a silicon interposer in between. It's not like you can pop it off and use it for something else.
Speculation: "AI Bankruptcy rigs" (Score:4, Interesting)
The sheer amount of not-quite-what-consumers-need RAM that is going to be auctioned after the AI companies' bankruptcies will be unprecedented. Remember we're literally talking about an amount equal to what the entire consumer market was using. It's going to be a few years worth of the entire RAM manufacturing industry!
The scale of not-quite-right RAM we're going to see, makes me imagine (warning: red flag for talking out my ass) consumer motherboards specifically made for the mountains of weirdly-packaged cheap-but-fast-as-fuck RAM at hyperscale data center auctions. We'd end up calling them "AI bankruptcy rigs" or something like that. Yes, they'll perform differently (more throughput than you need, more latency than you want), but it still ought to be usable. IGPs usually included, I suspect.
And they will just be a product of the times; ten years later no one would want one and bankruptcy rigs would be seen as odd late-2020s relics; the lemonade from Sam Altman's lemon industry. It would be a dead-end tech. But during those late 2020s? Could be a neat-o opportunity.
Re: (Score:3)
I expect to see a lot of GPUs with recycled RAM and maybe some recycled GPUs. You can already get such things from Chinese vendors who transplant parts from dead graphics cards, or repair them, or add video output to ones intended for compute only.
Re: No RAM, dwindling customers for AI (Score:5, Interesting)
Would be funny if "the people" raided Ai datacenters to get RAM. Your data center was raided by the the affordable Ram collective.
Re: No RAM, dwindling customers for AI (Score:2)
Then I'll know why I can no longer log on to the M$ account at work - someone raided the Microsoft datacenters
Re: (Score:2)
Would be funny if "the people" raided Ai datacenters to get RAM. Your data center was raided by the the affordable Ram collective.
I detect the seed of a South Park episode there. Maybe like this one. [wikipedia.org]
Re: (Score:3)
Some Chinese vendors do that on an industrial scale. If you look on AliExpress you will find things like graphics cards that use recycled mobile GPUs taken from dead laptops, and small form factor motherboards with recovered mobile CPUs on them.
Re:No RAM, dwindling customers for AI (Score:4, Informative)
The secondhand market is also a dumpster fire. Used hard drive prices are up and sellers are pulling ram from previous gen servers to gouge people.
Re: No RAM, dwindling customers for AI (Score:2)
Re: (Score:2)
Second, there are sites being built to increase capacity, but none dedicated to consumer.
Third, there is a bubble, and its about to pop. Unless governments keep stepping in to delay it....
Re: No RAM, dwindling customers for AI (Score:3)
Used DDR3 prices will increase
Re: (Score:2)
What do you put that into? Used MBs? What about the CPUs on those MBs?
I spent a month earlier this year scraping together a bunch of NOS servers, CPUs (early-gen Xeons) and DDR3 RAM for a legacy contract (with "frozen" software) that was nearing its term. It's tough to find and assemble 10 year-old tech at scale these days but it was kinda fun.
It's not a general solution for today's needs though.
nobody should ever need more then 640KB or RAM... (Score:2)
remember Bill Gates quote "nobody should ever need more then 640KB or RAM"?
geez that quote never happened. Why are people stupid enough to keep quoting something that was never said?
Re: (Score:2)
When we set the upper limit of PC-DOS at 640K, we thought nobody would ever need that much memory. -- William Gates, chairman of Microsoft
This quote appears at the beginning of an InfoWorld article by James Fawcette dated April 29, 1985. Fawcette gave no reference, and Gates denies saying it.
So, perhaps Gates never said it, but somebody claimed he said it, and that's how the urban legend began.
Re: nobody should ever need more then 640KB or RAM (Score:2)
Re: (Score:2)
The problem is that consumers aren't performance limited. For many years we have had over performing devices with hardware wasted on poor coding and stupidity. No you didn't need to upgrade your phone. Yes you could have paid someone to replace the battery, and to fix your 5 year old laptop keyboard. And both of those devices *can* run AI.
I'm starting to think about the positives here. With Microsoft forced to release lower RAM laptops they are starting to also release their shitty bloated OS is shitty and
Re: (Score:2)
there will be an ever dwindling number of consumer devices able to use AI services
Except AI isn't being developed for the use of consumers primarily. It's being developed to lower labour costs, facilitate disinformation, and for the implementation of mass surveillance of any imaginable social activity. Institutions will pay what's necessary for these capabilities.
Re: (Score:2)
An Example of Market Failure (Score:5, Insightful)
Re: (Score:3)
At this point computer storage is a national security issue and a huge economic problem for governments.
Re: An Example of Market Failure (Score:2)
International security issue.
Re:An Example of Market Failure (Score:4, Interesting)
Eh, market logic as designed (as high profits should attract more manufacturers).
But...
The amount of cash sloshing through will probably purchase that supply too.
The problem isn't exactly the market but the gradient of cash. When only dollars get a vote, the distribution starts looking tyrannical.
Re:An Example of Market Failure (Score:4, Insightful)
The real problem is that the companies soaking up the resources are doing it with OPM. Dry up the easy money and the egregiously high leverage and the RAM hoovering will subside.
Though I have to say the schadenfreude when Larry Ellison blows through most of his capital chasing this business by the tail is going to be somewhat satisfying to see. One good market correction and the street is going to aggressively make margin calls and investors are going to be so pre-occupied avoiding being the next Leopold Aschenbrenner that the "easy money" will dry up.
Stock up on popcorn.
Re: (Score:2)
Re: (Score:3)
Oh "the market" is not "working", this is a market failure. A working market requires the people within to act rationally and sensibly. But in this day and age, with LLMs and other generatives that are widely and accurately regarded as worthless, pointless, dismissible, futureless, counterproductive junk, it is clear that the tech pros and their investors are throwing good trillions after bad. Which would matter much, except that nearly everyone with any form of retirement savings or superannuation, is an investor in an index fund; is thereby an investor in much of this dipshittery, and is going to lose.some moola.
You are considering the current AI tech as "... junk". Many others aren't, that's why there is a gold rush - many see the value, and that it keeps on improving. Fast. Others again just go for the gold rush. Personally, I see it as a very valuable tool in my arsenal. It's not the only one, but it's a good, multi-purpose one. And it definitely needs to be checked, but the good ones give you references you can and should check when providing information - and for coding, it gives you what you ask for. Which h
Re: An Example of Market Failure (Score:3)
Re: (Score:2)
Markets need referees and we fired the ref. We didn't like bureaucrats and we didn't
Re: (Score:2)
This has been a slow moving train wreck for a year. Another year of production is fully sold out (likely the next 5). For someone to have not put plans in place last year to add capacity *somehow* is a market failure. The risk at current RAM prices is minimal, and dumping is impossible if inventory is sold out.
The key is to keep the additional capacity secret for as long as possible, until it is actually online.
Re: (Score:2)
The key is to keep the additional capacity secret for as long as possible, until it is actually online.
In many countries the local politicians would immediately announce and take credit for any such large investment in their jurisdiction (not to mention that any actual construction of such multi-billion dollar fabs is really hard for the community to miss (trucks, dirt, etc.)).
Re: An Example of Market Failure (Score:2)
Can they drop prices if the majority of their capacity is spoken for over the next five years?
Re: (Score:2)
Whether or not this qualifies as 'the market is working' is another question. I'd put it as another example where leaving things to the market does not serve people's needs.
It's not a failure for the market side of the equation, because we're captive. It's a failure for the consumer side. Alas, despite understanding that they are consumers, and that corporations are legal fictions whose very existence is allegedly in the interest of The People, there is very little will to force them to actually serve our interests before then going on to serve their own. This is in the name of "freedom" but somehow giving more rights to corporations doesn't seem to ever result in actually inc
Re: (Score:2)
The market optimising for profit and income in the short term does exactly this: AI companies are willing to pay more.
The market doesn't optimize for profit in the short term or in the long term.
In economics, there are two curves: a supply curve, and a demand curve. In a free market, they meet at equilibrium [wikipedia.org].
Re: (Score:2)
There has been no "demand curve" in the science of economics since the Arrow's theorem. The economics picture of "market equilibrium" that you were irradiated to simply doesn't exist.
Re: (Score:2)
You're ignoring the time lag.
Assuming your assumptions are correct, EVENTUALLY things would stabilize. But it can take multiple years to build a plant for making RAM. And it's a high risk investment, because while you're building it, other changes might happen.
Re: (Score:2)
You're ignoring the time lag.
Assuming your assumptions are correct, EVENTUALLY things would stabilize. But it can take multiple years to build a plant for making RAM. And it's a high risk investment, because while you're building it, other changes might happen.
You can argue that there is equilibrium right now - companies are selling and buying, it's just not at a price point consumers like. Thus we lower our demand.
Re: (Score:2)
Massively stupid (Score:5, Interesting)
They're just opening the doors wide for cheap Chinese memory, which will come, and then it will be RAMageddon for the cabal of manufacturers who so smugly announce that they've locked in obscene profits for years. With prices twice or more expensive than they need to be, there are many markets where it doesn't matter that China can't manufacture the same densities (yet). With enough money flowing in and this strategic opportunity given to them in a completely unforced fault, China will not take long enough to catch up for the established manufacturers to ramp up and do something about it.
Re:Massively stupid (Score:5, Interesting)
They're just opening the doors wide for cheap Chinese memory, which will come, ...
... and (assuming this happens prior to 2029) then it will get banned in America. But at least other countries will be able to buy it.
Re:Massively stupid (Score:5, Insightful)
Re: (Score:2)
Yep, let the rest of the world deal with RAM shitting itself in a short period of time due to shoddy and shady manufacturing. Open any computer system from about 20-25 years ago to see the leaking result of Chinese shit flooding the market.
Re: (Score:2)
Open any computer system from about 20-25 years ago to see the leaking result of Chinese shit flooding the market.
Good and bad parts alike are made in China. Nobody seems really clear on what exactly happened with the capacitors, which frankly makes me believe the corporate espionage story since it's the most nefarious presented so far and muddy waters primarily benefit those trying to hide something in the river. The scope of that particular problem hasn't repeated since, so clearly it was an exceptional circumstance.
Re: (Score:2)
Open any Chinese router from the past 15 years and look gloriously upon the shit called firmware in it. As a bonus you may find a door facing backwards! Go to Amazon and be amazed at the amount of Chinesium crap counterfeit goods masquerading as quality items. Trust in Chinese products is non-existent. The Chinese have a saying, "if you can cheat, then cheat". They take that to heart apparently.
Re:Massively stupid (Score:5, Insightful)
Chinese companies learn from the best (us). And they are only responding to demand. The blame for low-quality goods cannot be placed on China. The blame is squarely back here at home because low-quality, cheap goods are exactly what the American market is demanding, and demanding big time (Temu and Shein are wildly popular). No one wants high quality goods at higher cost. We all say we do, but in reality we don't, and our buying habits just reinforce that. No you can't blame China, sorry. The destruction of American industry is entirely a problem of our own system. American companies deliberately destroyed American manufacturing (and employment) in the name of short-term, quarterly profit. China didn't cheat and steal our jobs: we gave it all to them on purpose. No amount of tariffs and taxes will change that.
And for the record, I've had to deal with poor quality goods made in the US as well. Cutting corners to save pennies applies just as much to American manufacturers (maybe even more so) as it does to Chinese manufacturers. For example, Eaton Cutler-Hammer electrical panels are the most flimsy things I've ever dealt with. Every possible cost was cut.
Re:Massively stupid (Score:4, Insightful)
Well...that's not really accurate. Many people WANT higher quality goods, even at a higher price, but they can't tell what they'd be getting, so they opt for the lower price. Higher prices don't guarantee better quality, they just allow it. But when the market is opaque, all that's penalized is higher costs.
Re: (Score:2)
And for the record, I've had to deal with poor quality goods made in the US as well. Cutting corners to save pennies applies just as much to American manufacturers (maybe even more so) as it does to Chinese manufacturers. For example, Eaton Cutler-Hammer electrical panels are the most flimsy things I've ever dealt with. Every possible cost was cut.
FWIW, many of us outside the US don't think of the US as a country manufacturing high quality goods. For those, you'd go to Germany, Switzerland, Japan, or Taiwan. Or even China. The US is the country where anything goes. Cars being one example - clunky, low quality, low tech/gas guzzling, poor to drive vehicles of enormous sizes. A US company optimizes for low production quality vs. warranty costs, it doesn't optimize for high quality and see warranties as a failure.
Now, having lived and worked in the US
Re: (Score:2)
Chinese companies learn from the best (us). And they are only responding to demand. The blame for low-quality goods cannot be placed on China. The blame is squarely back here at home because low-quality, cheap goods are exactly what the American market is demanding, and demanding big time (Temu and Shein are wildly popular).
I'm old enough to remember when Japanese products had this exact same reputation. I'm really surprised at how short-term some people's memories are - especially a lot of our leaders, who are old enough to also remember it.
Re: (Score:2)
Re: (Score:3)
Firmware in most consumer routers is based on Linux, with the router vendor adding some UI and chip-vendor drivers to the open source baseline. Not written from scratch. And thus just as good or bad as routers whose hardware was assembled elsewhere.
China can't manufacture the high end chips yet, but they do have their own supply chain for midrange SOCs and memory. Gigadevices cloned some European SOCs. They work. Chinese contract manufacturers build to the quality level the customer pays for. You want cheap
Re:Massively stupid (Score:4, Informative)
CXMT is already making DDR5. It's not the fastest, but it works. Hopefully output can ramp up really fast.
Re: (Score:2)
Re: (Score:2)
Re: (Score:3)
I think you AI haters have some stupid ideas about what is going on.
The RAM is for data centers that host AI software. Selling the capacity to ordinary companies that use AI in their business processes.
For example: I host 25 agents on rented hardware in a data center.
As long as there are ordinary companies that reform their IT from "CRMs" to agent driven solutions: those new data centers will have customers.
And AI in "ordinary companies" like a vacuum machine seller, or a solar company: is not going away.
Th
Re: Massively stupid (Score:2)
It seems possible to me that they are overbuilding and over estimating demand increase.
Similar to how WorldCom did during the internet bubble.
AI isn't going anywhere, but that doesn't mean a lot of datacenters won't fail and get snatched up by companies that didn't (much like the WorldCom capacity).
Re: (Score:2)
They're just opening the doors wide for cheap Chinese memory, which will come
Except that's not how the semiconductor industry works at all. This isn't a labour intensive process that can be done cheaply in China. It's a highly specialised technical manufacturing process that is expensive regardless of where we are. We already have this, ChangXin Memory Technologies (CXMT if you haven't heard of em) is currently priced *HIGHER* that Samsung's for raw memory chips of the same performance.
And guess what, ... they are also sold out of capacity.
If the Chinese could have undercut their pe
So no more new data centers? (Score:2)
I just wish this whole AI as a service bubble would hurry up and deflate.
Re: So no more new data centers? (Score:3)
Re: (Score:2)
The Americans have made it illegal to possess more than 4 GB of RAM. They're setting up amnesty stations.
https://en.wikipedia.org/wiki/... [wikipedia.org]
Re: So no more new data centers? (Score:2)
According to that link it's spent about 1/3 to 1/2 the time less than now since 2011.
Additionally it was much less in 2017 (but had a peak higher in 2018).
This is also the largest and sharpest rise since 2010.
Moore's law finally defeated (Score:2)
Microsoft should be forced to extend Windows 10 until the RAM apocalypse is over or drop the artificial upgrade blocks to Windows 11. Yes Linux is an option for around ~5-10% of users now, too. Bu
Re: Moore's law finally defeated (Score:2)
Got my old Pentium 75MHz laptop out to show the kids last week. 40Mb of RAM, windows 98. Was surprised how smooth it ran. No USB, no internet but it has a floppy drive!
Re: (Score:2)
At work I run Windows 11 on 4GB. If I leave it running for more than a few days then office windows stop drawing correctly, and don't even get me started on Acrobat.
I own a 2C4T Zen+ laptop which came with 4GB, I was able to get a matching SODIMM on ebay for about $15 and it's also dual channel which is especially nice for a unified memory machine. It's running Devuan with KDE and it's absolutely fine for any task I'd want to do on it. But it would cover my needs for Windows, too. 4GB doesn't, and didn't wi
Encouraging MSFT to improve Windows is a bad idea. (Score:2)
When proprietary software makes FOSS more attractive that's to be encouraged so companies have more incentive not less to support Linux. The worse Windows gets the better.
Games are trifles compared to freedom and FOSS is critical to freedom of communication and computing now and in future. Pretending to be a cartoon character is fun but gaming is hardly critical to quality of life.
Re: (Score:2)
RAM is cheaper today than it was in 2017. Even without correcting for inflation.
https://dam.stanford.edu/memor... [stanford.edu]
or they have passed the same pot of on paper money (Score:2)
SW devs: RAM is no longer plenty and cheap (Score:2)
Re: SW devs: RAM is no longer plenty and cheap (Score:2)
Exactly, it's been too long where optimization has been ignored due to cheap and plentiful pricing. Apple is the exception as it has tight control over its ram usage and could ship with 8gb prior to AI. That was due to squeezing profit out, regardless of the reason, it can be done.
Re: (Score:2)
Whatever they're doing on the Neo front, it's been successful. I picked up several for my kids while they were at the $599 price point. They're plenty "good enough" and at least I know Apple will support them for a few years with OS updates/patches.
Re: (Score:2)
Scrutinize every instruction in assembly language, like the engineers did with the Apollo Guidance Computer.
This doesn't mean every line of source code in a high level language, they'll be analyzing what the compiler emits.
And now... (Score:4, Insightful)
...after that AI companies have been allowed to use your electricity, your water, and your land, they will be allowed to seize RAM from your laptop.
Re: (Score:2)
That's all under the assumption that the bubble... (Score:2)
... hasn't burst by 2027.
Considering that companies, the sources of income OpenAI and Anthropic hope to have, are now reacting to recent price hikes... by looking into cheaper "non-frontier" models... perhaps even on their own hardware, it may just be weeks or months until the bubble bursts.
Why spend 10x the money, to get an advantage you can neither measure nor feel?
It seems that with AI, the more effort you put into it, the less profit you get.
Re: (Score:2, Insightful)
There seriously isn't an AI bubble.
Re: (Score:2)
Do you have any kind of evidence to back up your claim? Sure there will be AI in areas where it's useful, but the current hype of generating ever larger and larger "frontier models" will end very soon, that is virtually certain.
Such models aren't much more useful than small ones, and once someone discovers a trick to make such small models even a bit better (e.g. run-time re-enforcement learning for the task you want to solve) there will no longer be the need for such huge data centers or companies like Ant
Re: (Score:2)
Says the Guy who knows nothing about AI.
The trend is going to multi model models. As in you have a few hundred small models interacting with each other (models, not agents).
At the moment we are at the level of Large Language Models. So the thing they do is in the name: they work on Language.
The next step is to work on Thinking. It is only a matter of a few years - perhaps three - and we have thinking machines. And that has absolutely nothing to do with RAM, or the lack thereof, but SOFTWARE.
Software as in m
Re: (Score:2)
Again, that doesn't contradict the argument that we are in a bubble. Actual thinking is something much different to LLMs and even if LLMs are used as part of such a system, they will be much smaller. Frontier-Level LLMs are _already_ to expensive to ever be economical, and bubbles are not about technology, but economics.
Re: (Score:2)
Why do you consider them expensive?
You can train one on your laptop for a few 100 dollars cost of electricity.
Expensive might have been the recent 30 years of research: and that is already paid for.
Re: (Score:2)
Yeah, but that's actually not our problem here. If everyone would just do that, we'd have no memory crisis. It's the "hyperscalers" who are gobbling up all that memory for their _huge_ models. That's the problem, not someone running something locally.
And companies are already switching to local models instead of paying those hyperscalers more money. That means those hyperscalers will loose more and more of their income. That will cause the bubble to burst.
Re: (Score:2)
IIUC, the evidence is that there are several different AI bubbles going on at the same time. They probably won't collapse at the same time, and even when the last collapsed AI will still be around. There are lots of jobs it's good enough at that it won't be replaced.
Yeah, there *are* AI bubbles. Groups of speculators are outbidding any reasonable return on several fragments of the market. But that doesn't make the actual tech go away, and there are domains were it's currently superior to any alternative
Can you re-use old DDR3 or DDR4? (Score:2)
I don't know why, but I've removed the RAM from every old computer that I've sent to recycling (after wiping the disks, of course).
I have boxes of DDR3 and DDR4 RAM. I think I can re-use the DDR4 stuff in some relatively modern stuff... don't know about DDR3...it might be too slow
Re: (Score:2)
There are some motherboards which will let you use DDR4 with an AM4 processor, and AM4 processors have made sort of a comeback with AMD bringing back production of some higher end parts which remain competitive.
Though I actually built this as the AM5 stuff drove down the prices rather than going back and doing it, I run a 5900X (12C24T 3.7-4.8 GHz) and it remains a beast given the costs. I use 4x CP16G4DFRA32A.C8FF for 64GB dual channel, and the mainboard is an ASRock B550M PG Riptide. It's been absolutely
Re: (Score:2)
Re: (Score:2)
Yes of course. You can buy brand new computers today that use either DDR3 or 4. Lots of places recommend going with DDR4 because the premium for 5 isn't worth it.
Reminder, there is the big 3, and the small 2 (Score:4, Informative)
On a positive note:
Aside from Micron, Samsung and SKHynix, that are able to porduce DDR 4&5, LPDDR 4&5, HBM and GDDR, there are Nanya (Japan ~3%) market share, and CXMT (China ~5% Market share).
Those companies only produce DDR 4&5 and LPDDR 4&5 (but CXMT in gearing up for HBM)
Further downmarket is Winbond (Taiwan), which produces LPDDR and DDR 3 & 4.
YMTC is gearing up to produce LPDDR5
But on a not so positive note, please notice that DDR5 can be used for "normal" servers (in the form of Regustered-FullyBuffered-ECC-RAM Sticks), while LPDDR5 can be used for servers and AI accelerators mainly in the form of SOCAMM2
Coupons..... (Score:2)
Ration Stamps (Score:2)
These were used during world war 2 in the united states.
Good (Score:2)
Re:If Only It Were Possible... (Score:4, Informative)
...to make new factories.
Like imagine if it were possible to create new semiconductor factories, making chips en masse, -- you could perhaps do it in a BIIIG state with a lot of land. You could have a factory with like 100 million square feet. Wouldn't that be something?
Maybe we should have more factories.
If only there were a price incentive to do so.
SK Hynix is spending $38 billion to build two factories [engadget.com]: one for RAM, one for NAND. They expect to start clean room operations near the middle of 2029 at the earliest.
Building the factories isn't the issue. It's all the equipment needed to make/print the wafers, the equipment needed to keep things running, and perhaps most importantly, the people needed to run the factory. None of these can be found on Craisglist.
Re: (Score:2)
Literally one company makes the lithography machines.
Re: (Score:2)
No, there are at least 10 and I think most of them have systems that support 30nm masks, which is typical for DDR4. I don't know too much about Chinese capabilities, but I would think they could probably do DDR4 masks these days.
Re: (Score:3)
Building the factories isn't the issue.
Well, actually the AI DC builds have turned building any new fab into an issue too. Lead times simply on the large scale transformers needed is now in excess of 18 months (up to 4 years), and the DC builds have been hiring the same construction firms and trades people (think electricians, plumbers, etc.) as the fabs need to build out. There is just so much construction infrastructure that is available (and it, too, cannot be built up quickly).
It's all the equipment needed to make/print the wafers, the equipment needed to keep things running, and perhaps most importantly, the people needed to run the factory. None of these can be found on Craisglist.
ASML now has a lead time of roughly 2 years for a new machin
Re: If Only It Were Possible... (Score:2)