331
RAM to see 'meaningful price increases' before 2026 is out, as the memory crisis is predicted to last 'at least' another 3 years (www.techradar.com)
submitted 16 hours ago by to c/technology@lemmy.world
117 comments fedilink hide all child comments
you are viewing a single comment's thread
view the rest of the comments
[–] 18 points 15 hours ago* (18 children)

Unfortunately, not really. RAM made as HBM or put in weird data center modules is stuck in that form factor, and when the bubble does pop, there is loads of pent up demand for memory makers to fill with new production.

It’s not like crypto where they used consumer stuff.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 18 child comments
    •
  • [–] 2 points 10 hours ago (5 children)

    I had a Radeon VII, and it has HBM. The VRAM was soooo fucking fast. So if the dirt cheap BRING IT ON.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [–] 0 points 10 hours ago* (4 children)

    Datacenter GPUs can't game :(. As in, it's literally impossible, as they are missing ROPs.

    The last generation that could is basically the Nvidia V100, and the datacenter equivalent of your Radeon VII. AMD removed the ROPs after the Radeon VII generation. The RTX 3000-generation Nvidia A100 can technically run a game if hacked to do so, but it performs extremely poorly because of a lopsided compute config.

    Now, if you are interested in GPU compute and self-hosting, maybe we can use them.

    But Nvidia learned to buy back and destroy GPUs during the last crypto bust, so I'm afraid they may do it again.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 4 child comments
    •
  • [–] 2 points 9 hours ago (3 children)

    Sorry I wasnt taking about Datacentre GPUs. I was talking about HBM going unused.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 0 points 9 hours ago* (2 children)

    I know.

    One can't just bolt HBM onto something else, no matter how skilled the hacker is. It needs an interposer to connect it because the pin pitch is so small, and it needs a memory controller that supports it.

    EDIT: And if you're talking about AMD using it in gaming GPUs or some other hardware, unfortunately the design pipeline is very long. If AMD decided to make a gaming HBM GPU right this second, and rushed it, it would take years to design it and get it in production.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 2 points 11 hours ago (1 child)

    Just 3D print some adaptors.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 1 point 10 hours ago*

    If I had a few billion for a TSMC factory, I would...

    This is what I'm referencing. The memory sits on the package with the processor, connected with thousands of microscopic traces, so fine it needs an expensive substrate rather than a typical PCB, and expensive equipment to assemble:

    The connectors are tiny, and made with immense precision. There are about a thousand reasons you can't just transfer HBM once its been used, and besides; other memory controllers wouldnt' support it.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 6 points 15 hours ago (5 children)

    Also lots of people in threads like these are assuming that "the AI bubble pops" means "everybody stops using AI and it all goes back to the way things were before ChatGPT came along." There's still an enormous demand for AI inference, that's simply not going to go away. People will still want to use AI. The bubble popping would mean a big drop in resources being spent on developing new models, that's all. It'd help the prices somewhat in the short term. But in the long term we simply need more chip foundries.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [–] 1 point 7 hours ago (1 child)

    It's more than just that. Smaller models are improving rapidly and are already "good enough" for the majority of consumer use cases. It is entirely realistic that in a few years, there will be virtually no demand for trillion-parameter LLMs. Demand for inference might not shrink in terms of functionality, but it will absolutely shrink in terms of memory and compute requirements. And then there will be a hell of a lot more server capacity available than anyone needs or wants, because they're over-investing like crazy now.

    What happens when hundreds of billions of dollars spent on datacenters basically goes *poof*?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 1 point 6 hours ago

    The people who paid for the data centers will go bankrupt, the data centers themselves will be sold for pennies on the dollar, but then the people who bought those data centers for pennies on the dollar will be able to make profit renting inference for a lower price than those original investors could have sustained themselves on. So I'm expecting there'll still be plenty of AI horsepower churning away, it'll just be doing it more cheaply and not doing it under the banners of the current incumbents.

    First movers often fail in this manner, they spend a lot of money making mistakes and discoveries that later followers can make use of for cheap.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 4 points 14 hours ago (1 child)

    Never underestimate a nerd and a need to compute. Those modules will have someone building adapters and interfaces to use that shit. As soon as someone runs doom on one its game on.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 8 points 14 hours ago* (last edited 14 hours ago)

    HBM is only usable with chips co-designed for it, packaged on an interposer with microscopic traces. An enthusiast is not doing anything useful with it unless they own a few TSMC assembly plants.

    The CPU modules can be tricky, too.

    Now, some enthusiasts are already hacking. Go to the Level 1 Tech forums, and you’ll find a group trying to reverse engineer AMD Instinct GPUs designed for server motherboards, dealing with locked firmware, soldering stuff, you name it.

    …But you probably won’t like the end goal of the hacking: using them to run LLMs and other GPGPU stuff locally. They aren’t usable for gaming.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 1 point 14 hours ago (1 child)

    Good point, still there will be a massive amount of production capacity, and not enough customers to fill that capacity, so the prices will drop fast, as competition grows to utilize as much capacity as possible, because it's extremely expensive to shut down factories, or run factories at below capacity.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 1 point 10 hours ago*

    Unfortunately, memory production capacity changes vary slowly.

    In spite of the drama, memory production hasn't increased all that much, which is why there is such a RAM crisis. Maybe it will grow some before the bubble pops, but Micron, SK Hynix, and Samsung aren't stupid:

    https://en.sedaily.com/finance/2026/10/01/microns-long-term-chip-deals-jump-63-percent-as-ai-memory

    They are getting companies to sign long term (5+ year) pricing contracts. In other words, they clearly know the bubble could pop. They are already hedging their bets, and they aren't going to put themselves in the position of massive overexpansion.

  • permalink
  • fedilink
  • source
  • parent