Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

2027 memory capacity is reportedly sold out

ign.com

Post reply on HN
Print view View original

511–519 of 519 posts

Re: 2027 memory capacity is reportedly sold out

#511

Post by tjwebbnorfolk »

rickydroll wrote: post #479
tjwebbnorfolk wrote: post #199

Earlier quoted context omitted.

Inefficient per watt, yes, but local inference capacity is greatly underutilized in aggregate. If a model can run on a machine that already exists, that's a bunch of additional chips that don't need to be built.

You picked up on what I didn't write existing machines don't have to be built, you just use them. It occurred to me much later that local machines could become a shared resource of a small co-op. Or it only runs off solar or energy harvested at times of excess generation (too much daytime solar, lots of wind, or overnight when demand naturally drops). Exploring how to partition inference across many machines and shed…

yea right now local models are like the swing producer. When claude throws 503s or 429s at me, or I run out of quota, I fire up a local model.

It's also resilience. I can still do AI stuff if the internet goes down.

Re: 2027 memory capacity is reportedly sold out

#512

Post by gaigalas »

phire wrote: post #498
gaigalas wrote: post #399

Earlier quoted context omitted.

That is somehow inaccurate. DDR5 has significantly more challenging requirements. The inspection and quality control equipment is one of those that you can't just reuse without careful transitioning. DDR5 has much higher speeds, on-die ECC and all sorts of complications DDR4 doesn't have.

I'm not even sure inspection and quality control is done at the fab. At least with logic ICs, they either testing/QC/bining as part of packaging (before dicing), or even have factories that specialise in just testing/QC/Binning. Such fallacies need to be flexible enough to deal with thousands of different IC designs, and there is no reason to tie them to a given process node. I guess with memory, there are only three…

> packaging (before dicing)

packaging before dicing? That makes no sense.

Re: 2027 memory capacity is reportedly sold out

#513

Post by gaigalas »

zahlman wrote: post #500
gaigalas wrote: post #396

Earlier quoted context omitted.

My comment was in the context of fabrication and how that affects the price of DDR4. DDR5 fabrication requires a sub 15mm process, much more complex mounting and lots more testing equipment. It means you can't just quickly convert a DDR4 fab to DDR5. It was meant to explain that DDR4 price increase has nothing to do with supply chain lanes, and it's more of an artificial demand manipulation than anything else.

>explain that DDR4 price increase has nothing to do with supply chain lanes, and it's more of an artificial demand manipulation There's no evidence of anything artificial here, except to the extent that the data center-driven demand could be called "artificial". When prices for DDR5 go up, it's only to be expected that this induces demand for DDR4, the spillover from people who were priced out of DDR5. Much the same…

> There's no evidence of anything artificial here, except to the extent that the data center-driven demand could be called "artificial"

I'm sorry. Can you summarize that part in a more concise phrase?

Re: 2027 memory capacity is reportedly sold out

#515

Post by jquery »

clircle wrote: post #369
jquery wrote: post #308

Earlier quoted context omitted.

Reading this thread I am so frickin glad I built a monster DDR4 workstation back in 2021. It would cost me more to build out that computer today (dual 3090s, 512GB ECC, 3995wx). Ram was so cheap back then.

Very neat. What's the power draw on this machine?

From the wall: 260W idle, ~1200W load (1600W PSU)

Re: 2027 memory capacity is reportedly sold out

#516

Post by phire »

gaigalas wrote: post #512
phire wrote: post #498

Earlier quoted context omitted.

I'm not even sure inspection and quality control is done at the fab. At least with logic ICs, they either testing/QC/bining as part of packaging (before dicing), or even have factories that specialise in just testing/QC/Binning. Such fallacies need to be flexible enough to deal with thousands of different IC designs, and there is no reason to tie them to a given process node. I guess with memory, there are only three…

> packaging (before dicing) packaging before dicing? That makes no sense.

IMO, dicing counts as part of packaging. So before the dicing step of packaging.

Re: 2027 memory capacity is reportedly sold out

#517

Post by gaigalas »

phire wrote: post #516
gaigalas wrote: post #512

Earlier quoted context omitted.

> packaging (before dicing) packaging before dicing? That makes no sense.

IMO, dicing counts as part of packaging. So before the dicing step of packaging.

Each die on the wafer is kind of different. Light is not all the same, usually dies on the center of the wafer are better and it gets worse as you move outside.

That plays out in binning, and it involves even design (laying out dies on the wafer such as the imperfections produce still usable ones, perhaps with some stuff disabled or running on lower specs).

So, there is some sort of QA at that point (seeing how bad the bad dies are), which also feeds into adjusting following batches.

DDR5 is already more complicated at that step. As I mentioned, it has ECC on the die and all other sorts of complications. A clean reuse of a DDR4 process seems unfeasible. There's significant investment needed for DDR5 even for what you consider "dicing".

Re: 2027 memory capacity is reportedly sold out

#518

Post by Anamon »

hardbass wrote: post #499
noduerme wrote: post #328

So LLM profiteers have bought up all the consumer RAM for the next several years in order to deliver more of a service we're already oversaturated with that doesn't turn a profit..? I guess by then there'll be no consumers left with jobs to afford a PlayStation 6, so no problem. Except how will the AI trainers find revenue? If maximum potential is reached by laying off 90% of workers at companies that sell stuff... h…

Who is holding the hands of RAM fabs and stopping them from making more? Its not the first time they have been caught doing collusion and price fixing tactics.

A more benevolent interpretation might be that they, too, see that this AI bubble will crash on the near future, and they'd be sitting on expensively built-out capacity that they don't need.

Re: 2027 memory capacity is reportedly sold out

#519

Post by Anamon »

1718627440 wrote: post #384
apple4ever wrote: post #92

This is why I'm so hesitant to use AI. There are other reasons (for example: the journey is the fun part, not the destination; and I want to learn not have something done for me), but the pressure AI puts on memory and storage is too much for me. I do use it occasionally, mostly to keep tabs on it and for small problem solving (how do I do this one piece of a query in MySQL, since I'm not a MySQL DBA), but that's it…

First, I am kinda the same. > how do I do this one piece of a query in MySQL, since I'm not a MySQL DBA Honestly for most things, the actual documentation is the best resource. I needed to learn to stop looking for solutions to my issue on various fora (like SO) and just read the primary documentation from the vendor. Have a look at dev.mysql.com or mariadb.com/docs/server

One of the problems with this is that documentation is increasingly also written with LLMs by the vendors, and now is often so vague and riddled with errors to be basically unusable, if not counterproductive by sending you the wrong way.

Microsoft has a great racket going on there. Vibecode shitty, unreliable services, offer wrong information in the slop documentation, then charge an arm and a leg for premium support which, in many cases, will just be an employee regurgitating more wrong AI slop information to you.

Post reply on HN