SK hynix dethrones Samsung as world’s top DRAM maker
186 points
by ksec
6 days ago
| 8 comments
| koreajoongangdaily.joins.com
| HN
ajb
1 day ago
[-]
I think it was Sophie Wilson who described the DRAM market as "a game of chicken with billions of dollars". Basically the market flip flops between glut and scarcity. When there are too many fabs, there is a glut and all the manufacturers grit their teeth and lose money, until one chickens out and closes a fab. Then prices go up and the ones still in the game make money, hand over fist, until too many fabs are built for the next generation of chips and the market flips back.

I haven't personally checked this against market data, FWIW.

reply
ksec
1 day ago
[-]
That is correct. Even when we are only left with three DRAM foundry, all three were watching out whether one of them will build a new Fab. That has created an environment where no one wants to build it first unless they are absolutely sure, and when one will be built all other follows.

And that is the reason why TSMC said they will never go into the DRAM business.

reply
stogot
2 days ago
[-]
I keep seeing this SK company’s brands everywhere. It’s a huge conglomerate, privately owned, that seems to be expanding rapidly and doing very well. Does anyone know why they’re so successful or is my misperception?
reply
dv_dt
2 days ago
[-]
Imho, if they're well managed, any private enterprise that is capitalized and has a long term outlook can run circles around public company management that can barely keep a year of marketing strategy consistent, let alone deep technical development.
reply
ahartmetz
2 days ago
[-]
Though, AFAICT, US public companies are the worst regarding financial short-term thinking. It's not as bad in other countries.

It is also possible to make the same mistakes for different reasons: lack of imagination, conservatism, entrenched interests...

reply
mathverse
2 days ago
[-]
Korean conglomerates are focused on hard sciences and there is very little room to do anything else that's why they excel here. They hire from the best Ivy style schools in Korea and focus on cutting edge stuff or improving cutting edge manufacturing

Here's the catch. Because of these constraints Korean conglomerates dont create as many jobs.Korean software or services industry is almost non existent or heavily constrained to Korea.

reply
marcyb5st
2 days ago
[-]
There is also to say that Chaebols (i.e. Korean conglomerates) can do basically whatever they want because the Korean government will bail them out/give them the financial support they need.

With this, the "I need to be extremely profitable" burden is somewhat lifted, giving them the freedom to do hard R&D.

And it is true now still, with the last bribery in exchange of favors dating back to 2018 [1]

[1] https://bruinpoliticalreview.org/articles?post-slug=south-ko...

reply
jjani
9 hours ago
[-]
> Here's the catch. Because of these constraints Korean conglomerates don't create as many jobs.Korean software or services industry is almost non existent or heavily constrained to Korea.

What? Korea's software jobs per capita is one of the highest of all wealthy (let's say top 30 HDI) countries. Please stop claiming this sort of stuff without being familiar with the country.

reply
WaxProlix
2 days ago
[-]
They've been around in one form or another since the early 80s, and have been in and out of a few of the major chaebols in that time. I wouldn't call them privately owned in the western sense, necessarily. There's no institutional investment behind the chaebols themselves generally, they're 'owned' by a single family and passed down hereditarily, and are nationalistic in a certain sense; they're much more closely ingrained with both government and state identity than most western corps.
reply
nerderloo
2 days ago
[-]
They're not privately owned. Bunch of SK companies are publicly traded in Korean stock market, including fore mentioned SK Hynix KRX: 000660
reply
butifnot0701
2 days ago
[-]
As others have noted, SK is not private - although it's hard to say whether it's private or public as it's more of group of companies. But many of larger companies including Holding Co is private. The company in this article's context is SK Hynix which is second largest (or third?) on Korean exchange. Just like other conglomerates in this country, SK Group runs many other businesses including bio, finance, telecom, etc.
reply
b473a
2 days ago
[-]
They purchased Intel's entire NAND business a few years back, so they just kind of exploded into the SSD market. They sometimes sell their drives under the name Solidigm.
reply
wtallis
2 days ago
[-]
I think SK Hynix's NAND business may have already been bigger than Intel's NAND business when they made that acquisition. Certainly by then SK Hynix had recovered from being late to the 3D NAND transition, while Intel was on a worse technological trajectory with their roadmap that diverged from the rest of the industry.

Your impression that they were at all new to the SSD market is largely due to the fact that SK Hynix operated mainly as a component supplier, and has never pursued promotion of their own retail SSD brand the way Samsung does. Hynix was a major player in the NAND industry before the SSD market as we know it even existed, and has been a major supplier of SSDs to PC OEMs for as long as PC OEMs have been buying SSDs in large volumes.

reply
seanmcdirmid
2 days ago
[-]
Aren’t all South Korean conglomerates like that? Samsung, SK, LG, Lotte, … they have shocking broad business lines even if you just know them for something more narrow.
reply
neom
2 days ago
[-]
Yes. You can technically drive Samsung branded cars, and live in LG built condo buildings.
reply
seanmcdirmid
1 day ago
[-]
Samsung makes container ships, they built the Burj Khalifa, they operate a theme park.
reply
vel0city
2 days ago
[-]
SK Group is one of the big Korean chaebols. An important thing to understand when looking at South Korean politics and economics.

https://en.m.wikipedia.org/wiki/Chaebol

reply
userbinator
2 days ago
[-]
https://en.wikipedia.org/wiki/SK_Group

Note that "SK" does not stand for "South Korea", as one might be lead to believe.

reply
numpad0
2 days ago
[-]
Hasn't this been a common trope for past ~50 years? East Asian anything is at least within a few miles to American/European anything, but way cheaper thanks to USD dominance, and that situation renders non-Asian industrial fabrication pointless and unsustainable, and East Asian products win.

From Toyota cars to Sony TVs to TSMC chips to DJI drones. It's been that way for a while.

reply
somat
2 days ago
[-]
So who makes the best ram? or is it largely interchangeable.

Off the top of my head there is only like three manufacturers left. Micron being the only one not mentioned here.

reply
wtallis
2 days ago
[-]
Samsung, SK Hynix, and Micron are the only three manufacturers doing leading-edge DRAM manufacturing, but there are other companies like Nanya that do older DRAM types and specialized DRAM parts for eg. embedded and industrial applications rather than PC, servers, and phones.

As for who makes the best RAM, it changes from one generation to the next, and also depends on what you consider "best": you might be looking for chips that overclock well in a desktop, or that are least likely to suffer compatibility issues and performance loss when maxing out the capacity in a desktop, or maybe "best" might mean who has the lowest-power LPDDR for your phone.

The DRAM parts made by the big three largely all adhere to the same standards (thought not necessarily all supporting the same frequencies), with the most significant recent example being GDDR6X that was essentially a NVIDIA-Micron exclusive partnership. For the most part, it's the latest iteration of DDR (desktops and servers), LPDDR (handhelds and low-power laptops), GDDR (GPUs), and HBM (more expensive GPUs).

reply
dehrmann
2 days ago
[-]
> As for who makes the best RAM, it changes from one generation to the next

This reminds me of HDDs and SSDs, though I've always found RAM to be generally reliable or obviously bad, while storage can look ok for a while before it fails.

reply
cm2187
2 days ago
[-]
And Samsung had a good reputation for enterprise SSD but obliterated it with pretty major firmware bugs (as in self destructing bugs, which they won’t patch without a support contract). SK Hynix seems to be a smaller player in that space.
reply
Cyph0n
1 day ago
[-]
HBM is also used in NPUs (network processing units) to be able to buffer and process packets at line rate (e.g., 400 Gbps) in high end switches and routers.
reply
ksec
2 days ago
[-]
Just to add CXMT from China has been catching up as well. Although I don't believe they will be up to leading edge DRAM given the current state of things.
reply
Panzer04
1 day ago
[-]
SK Hynix is winning in DDR5 land - their memory performs significantly better in all respects compared to Micron and Samsung.

Whether that matters much is debateable - maybe they get higher yields as a result (since more chips are sufficiently performant to be useful) but JEDEC specs seem pretty generous relative to what you can achieve on consumer platforms, so I somewhat doubt much RAM is thrown out because it's too slow to meet spec.

It's pretty random, though. Back in DDR4 days Samsung produced the best memory (B-die) and it wasn't particularly close - near-DDR5 speeds with lower latencies. At the same time, some of their other dies (I guess from other fabs?) was absolutely awful.

Again, I don't know how much that translates into profit though, since the performance user market for RAM is probably a fraction of a fraction of the overall memory market.

reply
pirocks
2 days ago
[-]
My sibling comment has a lot of good info, but to answer your question for ddr5, SK Hynix makes the ram capable of the highest frequencies/best timings, Samsung is second and micron is third. At least this was true when ddr5 initially came out, these things could change. Its also worth mentioning that manufactures can have different grades/product lines so there is variation within a manufacturer to. My perception is that Micron and Samsung are cheaper than SK Hynix, though idk what wholesale prices actually look like.
reply
Strom
2 days ago
[-]
In addition to generation, there are also differences based on capacity. I know less about DDR5 specifics, but with DDR4 you had Micron having the best chips for 32GB dual-rank sticks, while Samsung had the best for smaller capacities.
reply
noboostforyou
2 days ago
[-]
The "best" depends on what you want. For example, with desktop memory the previous DDR4 gen Samsung's B-die was considered the top choice if you wanted the fastest speeds and highest overclocking potential. Micron's rev-e die was also desirable as it was usually dual rank (technically slower, but for certain real world applications you will get better performance) and also overclocked quite well.

Now with DDR5, Hynix's A-die is considered the best option.

reply
Panzer04
1 day ago
[-]
Dual rank was a lot faster in DDR4 days IIRC, because it allowed a memory channel to have more requests in-flight on the memory - it could be worth like +10% performance or more.

Less impactful in DDR5 because they made it a design goal to have more "ranks" by default, though it does still have a small performance benefit.

reply
thehumanmeat
2 days ago
[-]
Slightly related, I've been wanting to invest in SK Hynix since early '23. Does anyone know how a US investor can get access to the stock? I haven't found any brokers that allow you to trade Korea Exchange tickers.
reply
cherryteastain
2 days ago
[-]
Interactive Brokers lets you trade HY9H and HXSCL which are secondary listings of SK Hynix on German and US exchanges respectively.
reply
neom
2 days ago
[-]
Wonder if the reserved tables order in Apgujeong shifts too.
reply
jacob_rezi
2 days ago
[-]
I still reserve a couple every month. I live 2 blocks away
reply
neom
2 days ago
[-]
I was referring to the fact that in Cheongdam many of the company man restaurants and clubs have a pecking order. Chey Tae-won now gets a better table? I wonder. :)
reply
mgh2
2 days ago
[-]
reply
downrightmike
2 days ago
[-]
Can we just make 64GB the minimum now please
reply
xiconfjs
2 days ago
[-]
Not for low power applications
reply
seanmcdirmid
2 days ago
[-]
64GB is what you need to run some decent quantized mid-sized LLMs locally…with unified memory on Apple silicon. Should be standard, that would open up a lot if new applications. Incidentally, even high DPI monitors aren’t standard yet for non-mobile devices. Sad how slowly things move.
reply
ZYbCRq22HbJ2y7
2 days ago
[-]
Are there aggregations of some accessible telemetry from a widely used application that reveal what is most common today?
reply
bcraven
2 days ago
[-]
https://store.steampowered.com/hwsurvey/Steam-Hardware-Softw...

Gamers only, but that's not a bad selection imho

reply
wtallis
2 days ago
[-]
The Steam Hardware Survey is an incredibly valuable resource, what with it being freely-available, constantly updated, and sourced from a population that makes its sampling biases generally easy to identify and understand. It and the Backblaze hard drive data are almost unique in how they provide real, large-scale data about computer hardware.
reply
Panzer04
1 day ago
[-]
It's interesting how you can practically tell how many people have a specific processor nowadays (12 core = 5900,7900,9900x, 14 core = i5 14600k, 245, 20 core = i7 etc)

Pretty cool

reply
wiredpancake
2 days ago
[-]
It has his flaws though and is somewhat archaic. I emailed Gabe about this a while back but got no response.
reply
danparsonson
2 days ago
[-]
Gosh he must've been busy with something
reply
Aleklart
2 days ago
[-]
No it is not, it just shows what Chinese bot farm emulators are configured at the moment.
reply
SweetSoftPillow
2 days ago
[-]
Why would any bot farm complete this survey?
reply
esseph
2 days ago
[-]
They're not, it's pulled from the application.
reply
wtallis
2 days ago
[-]
The user (or bot) doesn't have to provide the hardware information, but does have to provide consent for that information to be sent to Valve. Bots have no strong reason to prefer "yes" or "no", but if you have to implement code to handle that dialog box, answering "yes" might be seen as the more human-like behavior.
reply
ZYbCRq22HbJ2y7
2 days ago
[-]
Thanks, 35.15% on 32GB as well, getting there
reply
Brybry
2 days ago
[-]
I believe most consumer CPUs only have 2 memory channels w/ 1 memory controller so unless they're using 64GB UDIMMs (which I believe do exist as of this year) then gamers seem limited to 64GB total ram (2x32GB) unless they want to drop their ram frequency.

For example a 9950x3d officially supports 2 sticks at DDR5-5600 but 4 sticks at only DDR5-3600. [1]

I had a friend run into this issue on AM5 when he was trying to use 4x32GB DDR5 on his gaming PC.

[1] https://www.amd.com/en/products/processors/desktops/ryzen/90...

reply
wtallis
2 days ago
[-]
There have been 48GB dual-rank DIMMs for about two years now, so 96GB using two slots and operating at high frequency has been an option. But even 64GB is still somewhat overkill for a gaming PC, putting you more into workstation territory.
reply
numpad0
2 days ago
[-]
2x48GB is also like $250. Cheaper than weekly error margins for high end GPUs. It just don't make sense not to max out. Felt smoothness in OS, likely from disk cache, is also noticeable when "extra" capacity is removed.
reply
cute_boi
2 days ago
[-]
I asked my company to give me 32 GB RAM, then old boomer said why I need so much of RAM. They were asking whether I am building a rocket....
reply
jlokier
2 days ago
[-]
I've had that happen twice!

Recently I asked for my software developer colleage to be bought a 24 GB Macbook Air instead of 16 GB, and boss came back with "not everyone needs a super-big machine like yours Jamie!".

They seriously spent contractor time investigating whether 16 GB was "enough" to get by for our app development, for a price difference on one laptop (second hand) that was negligible compared with cost of my colleage's time.

When I was using 16 GB I regularly had to watch the spinning beachball waiting for tasks due to memory pressure. Between browsers and VMs, it was nowhere near enough for how I worked. So I knew why I was asking, and I knew the price difference was so small for the company, that it was a no-brainer. I gave justifications but it was seen as over-indulgent.

reply
resonious
2 days ago
[-]
I mean I also feel like 16GB should be more than enough for what I do (web dev). But bottom line is it isn't. I guess the people making these decisions should try building and running their app locally themselves...
reply
dlcarrier
1 day ago
[-]
A rocket can run on 32 KB of RAM. It's web browsers that take all the RAM. (At least for normal users, that don't run neural networks locally.)
reply
GloriousKoji
1 day ago
[-]
16GB has been the new minimum for under a year. Give it some more time /s
reply
walterbell
2 days ago
[-]
> This article was originally written in Korean and translated by a bilingual reporter with the help of generative AI tools. It was then edited by a native English-speaking editor. All AI-assisted translations are reviewed and refined by our newsroom.

Helpful footnote on man-machine boundary.

reply
jchw
2 days ago
[-]
Tangentially: LLMs are really impressive at translation. I guess it shouldn't come as that much of a surprise given where a lot of the most pivotal research came from, but still, the leading edge LLMs are extremely good for situations where having a human translator is infeasible or too expensive, and if you're worried about correctness you can go through and verify the translation using reference material and asking the LLM for more information about a given excerpt, which you can also verify against references and online discussions.

I think my only concern is that I'm not sure how to make sure I'll always have an untainted set of reference material to check against in the post-LLM Internet. We've had LLM hallucinations result in software features. Are we possibly headed towards a world where LLM hallucinations occasionally reshape language and slang?

I feel bad for human translators right now. For various use cases, current-day machine translations and especially LLM translations are sufficient. For those not versed in the world of otaku and video game nerds, one extremely fascinating development of the last few years is the one-shot commission platform Skeb, where people can send various kinds of art commission requests to Japanese artists. They integrate with DeepL to support requests from people who don't speak Japanese fluently, and it seems to generally work very well. (The lower-stakes nature of one-shot art commissions helps a bit here too, but at the least I think communication issues are rarely a huge problem, which is pretty impressive.) And that kicked off before LLMs started to push machine translations even further.

reply
pjc50
2 days ago
[-]
My wife has been working on translation recently, and the LLM hit rate for novels is highly variable. It's capable of just dropping out entire paragraphs. You still need a final pass from a human native speaker editor to check that it makes sense. Which is what's happening in this article, the news site cares enough about their brand and quality to check the output.

I agree that bidirectional communication is probably going to work a lot better, because people are more likely to be alert to the possibility of translation issues and can confirm understanding interactively.

reply
jchw
2 days ago
[-]
Definitely contextually dependent: things like news articles are probably the gold standard for machine translation, but creative works and particularly dense reading material like novels seem like they will remain a tougher nut to crack for even LLMs. It's not hopeless, but it's definitely way too early for anyone to be firing all of their translators.

> It's capable of just dropping out entire paragraphs.

I suspect, though, that issues like this can be fixed by improving how we interface with the LLM for the purposes of translation. (Closed-loop systems that use full LLMs under the hood but output a translation directly as-if they are just translation models probably already have solved this kind of problem by structuring the prompt carefully and possibly incrementally.)

reply
orthoxerox
2 days ago
[-]
They are. For unlicensed fan translations of indie Japanese games the word "MTL" used to mean "unplayable translation quality" until maybe a year or two ago. Now ChatGPT can maintain enough context to translate the game mostly correctly. There are still cases when the names flip-flop between two plausible translations (e.g. Rina-Lina or "scroll of wisdom"-"sage scroll") or the gender is not inferred correctly, but they are rare enough that a single editor can crowdsource and apply the fixes. The prose itself is now finally legible and you don't feel like you're reading the clues to a cryptic crossword.
reply
seanmcdirmid
2 days ago
[-]
I was watching a graphic novel on YouTube yesterday that was translated from Japanese text into an English narration. It was weird, not perfect, probably copyright infringement, but pretty effective. I think it’s only a matter of time until we have real time local translator hardware that we can just plug in our ear when traveling, or heck, working in another country where you don’t speak the local language. Language barriers are going to fall quickly.
reply
pjc50
2 days ago
[-]
The auto-dubbing on youtube has the classic hallmark of an AI product: you can't turn it off easily.

(the audio track switcher, which will give you back the original audio, is not available on mobile. Fortunately if you use newpipe it is ..)

> real time local translator hardware that we can just plug in our ear when traveling

These are definitely already a thing, popular in Asian countries. https://www.aliexpress.com/item/1005008777097933.html

reply
xeonmc
2 days ago
[-]
They could even make it fish-shaped!
reply
whatevermom
2 days ago
[-]
If they don’t, I will.
reply
eurleif
2 days ago
[-]
It's no doubt possible to have translation with less lag on average than current tools like Google Translate's Conversation Mode, but truly real-time translation is impossible because of word order differences. You can't translate a word you haven't heard yet.
reply
jcattle
2 days ago
[-]
Reminds me of a letter to the Times regarding German word order (and the fact that the verb often comes last)

> Sir, Your reader's reference to German word order reminds me of a UN meeting at which I worked when the German delegate ranted for ages while all French eyes turned to the French interpreter booth. The interpreter witheringly interjected "j'attends le verbe".

reply
ptero
2 days ago
[-]
But truly realtime translation is rarely (never?) critical. For streamed media there is usually enough of a buffer to handle word order swaps without delay (although one can make contrived counter-examoles), and for human interactions that delay is likely less than normal human processing time.
reply
lukax
2 days ago
[-]
Have you seen Soniox? They support real-time translation (only speech to text translation for now).

https://soniox.com/

(disclaimer: I worked there)

reply
seanmcdirmid
2 days ago
[-]
You know someone will figure out how to do real time translation and use the same voice as the speaker, or if you have a room with more than speaker, rather than figure out who you are listening to, just do them all and remix according to how the input sources were mixed.
reply
junon
2 days ago
[-]
There was just a demo for this posted like a week or two ago between an English and Korean speaker, I'll see if I can find it and edit/comment again.
reply
pezezin
2 days ago
[-]
I wonder how well that will work given the differences in grammar and word order. I have seen a similar solution applied between English and Japanese, which has a similar grammar to Korean, and the results were... not impressive.
reply
a-french-anon
2 days ago
[-]
"Watching" a graphic novel??
reply
qwertox
2 days ago
[-]
> LLMs are really impressive at translation.

I love it how one can mix languages in one sentence. "Anyway, re-add Tiefkühl and tell me to just check what they have there as TK-Gemüse, and a note at the end that I should go to EDEKA someday this or next week"

reply
thebruce87m
2 days ago
[-]
That paragraph itself could be spat out by the LLM too. I’m not saying it is, but it’s a depressing thought.

Maybe in the future they will spit this out as a result of the training data today, e.g. this article.

reply
anon-3988
2 days ago
[-]
Does it matter? Lets say it is translated by a person. So what? Is he qualified? Does he know technology enough to translate the terminologies properly? How do I even verify this?
reply