CPU shortage is reportedly 'more acute' than memory but shorter, as the industry banks on Intel 18A
Posted by sr_local@reddit | hardware | View on Reddit | 86 comments
TL;DR: The global CPU shortage is more severe but expected to be shorter than the ongoing DRAM crisis, which may last until 2030. Intel's ramp-up of its 18A process node with Panther Lake CPUs aims to ease supply issues, though reliance on TSMC remains critical for components and overall industry recovery.
Jmich96@reddit
I really don't understand the alleged CPU shortage. It's not as though AI data centers are going to buy 9600X or 14600kf CPUs for their server racks. Are companies like Intel and AMD shifting production to more valuable Xeon and EPYC CPUs?
My nearest Microcenter has full stock of even high-end consumer CPUS like the 9950X3D and even the 9970X.
ThePandaRider@reddit
Not sure what the root cause is but I do know that AI is being used to write a ton of poorly written and poorly optimized software.
pemb@reddit
This has been going on for a long, long time. Compilers were once maligned for enabling poorly written FORTRAN to be turned into poorly optimized assembly. Every innovation that lowers the skill and effort barrier to generating usable software will by definition enable lower skill/effort projects to exist, while also simultaneously pushing the limits of what is possible to achieve at the other end of the spectrum.
ThePandaRider@reddit
This one is slightly different. Before it was always a 1:1 relationship where one developer writes code and gets it to compile even if they are re-using libraries, now it's a one to many relationship where a single person spawns many agents to write requirements, come up with an implementation plan, write code, review, and adjust. If before you could get 10k lines of code out of a dev per month you can now get around 500k. The problem is that AI hallucinates and randomly invents requirements or writes something it shouldn't and then builds on top of that pretty quickly.
The point isn't the poor code quality. It's that there is a much greater quantity of code being pushed.
pemb@reddit
LOC is a poor measure of the amount of software being made, or programmer productivity in general. Always has been.
We're still in the first half of the adoption lifecycle of GenAI for software development, and it's not fundamentally different from the previous hot new tech, we're adding yet another abstraction layer between you and the logic gates and registers.
If you measure amount of software by lines of assembly, like our predecessors back in the 50s would, compilers also enabled an order of magnitude jump in output, but no one pays attention to that. If you measure it by binary size, it kept steadily bloating over the decades.
Uptons_BJs@reddit
Current generation Xeon (granite rapids) is fabbed on Intel 3 - Intel never built out a lot of capacity for Intel 3. Next generation Xeon is coming on 18A.
Current generation Epyc is fabbed on TSMC N4 and N3, and that is heavily capacity constrained.
Capital-Froyo-4359@reddit
Both the current and next generation Xeon processors use Intel 3. The new generation actually has more die area of Intel 3 than it does Intel 18a.
Xpander6@reddit
So why is there a shortage of intel 14th gen?
Geddagod@reddit
A huge mistake in hindsight, though I'm sure many people thought it was a mistake at the time too, and got steamrolled by Pat's enthusiasm about a fast and problem free 20A and 18A ramp.
Jmich96@reddit
This makes sense. However, at least on the consumer market, I'm not seeing a CPU shortage. Is this just data centers CPUs? If so, who cares (unless manufacturers shift to primarily data centers CPUs, like Nvidia with GPUs)?
Geddagod@reddit
Intel has stated they are deprioritizing the low end of client and shifting those Intel 7 wafers over to server.
jenny_905@reddit
Of course not but if the demand for server chips keeps going up then why would Intel or AMD devote extremely limited fab resources to producing the lower margin, lower demand consumer stuff?
It's the same issue as DRAM.
INITMalcanis@reddit
Once again, the inevitable outcome is that the consumer SKUs will be produced one process node further back from the best than at present.
makar1@reddit
Sales of CPUs are massively down due to RAM prices. CPUs are in stock because demand is even lower than supply.
red286@reddit
Yes. The issue is materials shortages, not CPUs being bought up. The problem is that the AI buildout is consuming all available semiconductors, which means that the cost to manufacture CPUs (and other components) is increasing. Value-focused processors will basically vanish because there's no way to maintain MSRP and make a profit, so they'll just quietly stop making them.
GenericUser1983@reddit
On the AMD side of things, the desktop processors use the same sort of chiplets that are also used in some of the server grade Epyc products. The desktop has 1 or 2 8 core (sometimes cut down to 6) chiplets attached to a basic i/o & igpu die, while some of the Epyc processors will have sixteen of the same chiplets attached to a much bigger i/o die. So increased datacenter demand could impact how many chiplets are available for desktop use. Could being a key word here; a lot of the desktop chiplets are actually ones that aren't good enough to be used in server stuff, and some server chips are using a different die (one with 12 of the smaller Zen 5c cores, made on the TSMC N3 process, not TSMC N4 like the other ones).
That said, I have noticed no shortage of CPUs either; plenty of them are regularly going on sale at various online retailers, and the local Best Buy & Walmart in my area have no problem keeping them in stock.
Geddagod@reddit
The N3E chiplet is 16 Zen 5C cores.
GenericUser1983@reddit
Thanks for the correction.
INITMalcanis@reddit
For God's sake, now it's CPUs as well? What next? Keyboards?
Capital-Froyo-4359@reddit
The ACTUAL biggest delays aren't in hardware at all.. they're in transformers, circuit breakers, gas turbines etc. Chips are worthless without power and these new AI chips consume insane amounts if power.
phylter99@reddit
I think it's all FUD and everyone knows if the news reports a shortage then the prices will go up and people will profit. They learned from the GPU shortage of a few years back.
Maimakterion@reddit
No it's not. LLM 'thinking' agents that run and uses normal software until a task finishes are now the standard. None of the software used by the LLM can run on solely on GPU arrays unlike old school LLM inference. Expensive inference GPU racks are going to idle if these providers can't get enough CPUs so they are willing to pay far far more than the list price.
fastheadcrab@reddit
AI agents have some potentially interesting applications but the hype around them is much more than the legit applications
imaginary_num6er@reddit
PSUs, CPU heatsinks, motherboard PCBs, copper in USB cables, etc.
nshire@reddit
I'm so tired of winning
BioshockEnthusiast@reddit
November can't come fast enough.
casvalniisan@reddit
Pray tell how would november change the literal shortages that's been going on for more than a year now
arcanemachined@reddit
Bad guy get lower number
INITMalcanis@reddit
I wish to return this future for a refund.
gahlo@reddit
We can give you store credit.
INITMalcanis@reddit
How much for another ten years of the 90s?
I liked the 90s.
gahlo@reddit
We're out of stock. We should be getting another shipment in 60 years or so.
INITMalcanis@reddit
Don't bother, RAM should be cheap again by then.
shpongolian@reddit
Those all use pretty standard materials that are used in many other industries, rather than specialized technology that requires billion dollar machines that only one or a handful of companies can produce
timfountain4444@reddit
And all the “pretty standard materials” use energy in their production and energy to move them from the production site to the consumer. And we have a global energy problem in case you missed it…
shpongolian@reddit
No shit but that applies to everything. Not remotely comparable to the RAM/CPU shortages
Saneless@reddit
No, not keyboards. Just the F and J keys that have the little nubs. Those are in short supply because the nub company can't handle the demand
Vb_33@reddit
Press F to show respects, oh wait...
gahlo@reddit
Good thing they aren't on the g key or they'd never be able to locate their stock.
kwirky88@reddit
Thanks, your humour made my day.
Saneless@reddit
Glad to hear it :)
kinkycarbon@reddit
It’s anything requiring silicon wafers.
dingo_xd@reddit
Be thankful that Chinese equipment for datacenters is also in shortage otherwise the situation could have been even worse for us.
wordswillneverhurtme@reddit
Imagine the bubble bursts. Oh boy how cheap the used market would be
Gwolf4@reddit
There is not going to be cheap hardware in this bubble. Data center gpu's don't use pcie consumer like we do. Some SSD need firmware to even boot. It is a disaster for the consumer.
It is likely they will be sold at lower price for second hand in smaller or lower priority industry.
No_Sheepherder_1855@reddit
They’re already capacity strained, even if the bubble bursts the need for current hardware will remain.
AnimalShithouse@reddit
The need for current hardware in a bubble burst scenario will nowhere near meet supply. Prices would come down overnight and then more of the "want" crowd would step in.
jenny_905@reddit
They're more likely to put unused stuff in the e-waste pile than sell it to plebs.
Arch-by-the-way@reddit
The bubble is already bursting but people won’t report it because their world view requires that they be a victim of corporate conspiracies
voiceipR@reddit
Another lie from Intel.
Baalii@reddit
Unless Intel suddenly books a ton of customers for 18A, it's safe to say that "the industry" disregards 18A and it just happens to come online during a shortage.
ConsistencyWelder@reddit
The yields of 18A are probably very low too. Low enough that they won't be able to supply enough to replace the current Arrow Lakes in their portfolio. I think we can expect to see 13th, 14th and 15th gen around for a looong time.
Saranhai@reddit
It's high enough for high volume manufacturing... Which generally is about 60% and Intel themselves say it's been improving month over month
Exist50@reddit
I think that's less about yield and more about a combination of performance and RnD effort.
CobaltFermi@reddit
Intel will need to ramp up 18A, not just for their own survival but also for the larger market to regain some of its stability. The article focuses on CPUs only but the bigger issue is that TSMC has a monopoly on the foundry business. Fingers crossed whether Intel Foundry can capture some of the massive demand from large customers like Microsoft, Nvidia, Apple.
Interesting-Rock2474@reddit
Which company besides Intel even uses 18a, Intel itself uses TSMC for high end products. If Intel ramps 18a harder it will have almost no impact. Samsung has a far larger impact and it has no need for Hype pieces written by ai or someone who has no understanding of the market.
Paed0philic_Jyu@reddit
18A offers what TSMC's advanced nodes don't - a node than can be exclusively used for CPUs depending on Intel's production capacity.
whatevermanbs@reddit
And the yields are coming only 2027ish .
Paed0philic_Jyu@reddit
Same as N2 for whoever is making datacenter CPUs on it then.
hwgod@reddit
Why would you compare 18A to N2? It's not competing in that class. And there are N3 server CPUs in production, unlike 18A.
Paed0philic_Jyu@reddit
CPUs on N3 are low volume relative to N5/N4 CPUs. The only x86 CPUs on it is Turin dense.
I'm not comparing 18A and N2. I'm commenting on when x86 server CPUs made on them are expected to have wide availability.
hwgod@reddit
You forgetting about ARL?
And why just x86? There's a ton of ARM chips using that node, and they have been for years.
Paed0philic_Jyu@reddit
What type of CPUs is the present discussion about? Consumer or datacenter?
Geddagod@reddit
This is the same dude who made the "N2 is a perf/watt and Fmax regression over N3" post on this subreddit a few months ago. On his like 3rd alt despite not caring about reddit apparently lol.
whatevermanbs@reddit
Can't even compared tsmc execution with intel's.intel date means date + 1 year.
Paed0philic_Jyu@reddit
Everything takes 1 year to reach acceptable volume supply. Both 18A and N2 datacenter x86 CPUs will be announced this year, but actual availability will be 2027.
whatevermanbs@reddit
You are running a little late on venice.
https://www.amd.com/en/newsroom/press-releases/2025-4-14-amd-achieves-first-tsmc-n2-product-silicon-milesto.html
Paed0philic_Jyu@reddit
That is just initial tapeout.
Are N2 datacenter products in the room with us right now?
whatevermanbs@reddit
Anyway.. you see the wafer picture of N2? So according to your calculation volume in 2026?
Paed0philic_Jyu@reddit
What do pictures of wafers have to do with when something would be widely available?
Besides, for Venice at least, AMD's bottleneck is TSMC's packaging, not N2.
AMD is totally dependent on how TSMC balances CoWoS and InFO capacity expansion for their upcoming products.
whatevermanbs@reddit
It is YOUR theory. Takes one year. It appears you just want to force venice to 27. So be it. But that would be 2 years from node announcement.
Paed0philic_Jyu@reddit
It is not a "theory". It is simply the likely outcome based on current timelines amidst supply constraints.
ThePandaRider@reddit
Intel still accounts for 70% of the server CPU market, 75% share of Laptop CPUs, and 63% of desktop CPUs as of the beginning of the year. Their next generation of server CPUs use 18a.
Exist50@reddit
But where are the wafers coming from? By volume, Intel's still heavily reliant on TSMC and older Intel nodes. GNR/SRF should be decently ramped by this point, but PTL is still early days, and it'll probably be another year-ish before DMR ramps.
ThePandaRider@reddit
Last year it was 70% inhouse and 30% from TSMC and expected to change to 80% in house with 18a coming online. They kept their server CPUs in-house but the high volume consumer CPUs were fabbed by TSMC for the last few generations. There is an estimate that 30k wafers per month came from TSMC in 2025. Fab 52 has a 40k wafer starts per month capacity.
What's being fabbed at TSMC for Diamond Rapids?
Exist50@reddit
A lot of the in house is from RPL/EMR on Intel 7. Curious what the breakdown would be with newer nodes.
At TSMC? Probably nothing, though they might be regretting that somewhat. But DMR is likely to have more Intel 3 than 18A silicon.
Exist50@reddit
Well, for some silicon.
whatevermanbs@reddit
How much of that 70% of server market is legacy node. And the big money is going to buy that shit? Well, they will if they have to. But it would be too stupid to think intel is going to make any decent money off it.
elkond@reddit
hey hey hey HEY
gtfo with facts, u r ruining the narrative!
gjt1337@reddit
You probably dont understand market. Maybe Intel isnt with cutting edge products but has capacity to meet increased global demand.
ammie12@reddit
even with new nodes ramping up supply diversification still feels pretty limited
Mrgluer@reddit
they’re playing factorio irl. ofc there’s gonna be bottlenecks that keep getting found.
Maimakterion@reddit
This CPU shortage is more intense because it turns out all the GPUs you bought for inference b/c you only needed 1 CPU per 8 GPU are going to go idle unless you scrounge up more CPUs.
III-V@reddit
Man, it's a breath of fresh air that Intel's moving on from 10nm. Intel 4/3 is kind of in a weird place in that they did so much life extension to 10nm that the node doesn't stand out, but from here on out, just having things be normal again... it'll be nice.
Obviously, the underlying economics and supply situation is a disaster, but I could never afford parts anyway, so I'm just happy I'll have stuff to read about again.
Geddagod@reddit
Is this not the same thing with every node after Intel 7 too?
Ofc the next generation nodes did always have their perf/watt and density uplifts, but Fmax was not great.
And while Intel 18A is expected to ship way more wafers than Intel 4/3, it's still not normal, since the compute tiles for NVL-S and the 4+8 tile in NVL-H are both rumored to be on N2 rather than 18A.
ConsistencyWelder@reddit
Now, until the AI crash happens, we might need to change the hardware we use as consumers to whatever the datacenters use. As they upgrade their hardware, huge amounts of RAM, SSD's, CPU's will become available on the used market. So let's take advantage of that and get some cheap ECC memory and a cheap Epyc CPU.