no comments (yet)

sorted by: hot top controversial new old
there doesn't seem to be anything here
57 comments in c/selfhosted@lemmy.world view that post
[–] 81 points 11 months ago (7 children)

Just buy an extra GPU and figure out how to turn it's VRAM into regular RAM. 😌

  • source
  • hideshow 7 child comments
  • [–] 36 points 11 months ago (4 children)

    I know you can create a gpu ram disk so you could use it as swap lol.

  • source
  • parent
  • hideshow 4 child comments
  • [–] 11 points 11 months ago (3 children)

    Ok, I have to know: how is this done, and what do people use it for?

  • source
  • parent
  • hideshow 3 child comments
  • [–] 14 points 11 months ago (2 children)

    It uses a github repo that hasn't been updated in 3 years and it's more of a gimmick but it's a fun weekend project.

    https://www.youtube.com/watch?v=qyzBRsfQ_UM

  • source
  • parent
  • hideshow 2 child comments
  • [–] 8 points 11 months ago (1 child)

    https://github.com/Overv/vramfs

    Oh, it's a user space (FUSE) driver. I was rather hoping it was an out-of-tree Linux kernel driver, since using FUSE will: 1) always pass back to userspace, which costs performance, and 2) destroys any possibility of DMA-enabled memory operations (DPDK is a possible exception). I suppose if the only objective was to store files in VRAM, this does technically meet that, but it's leaving quite a lot on the table, IMO.

    If this were a kernel module, the filesystem performance would presumably improve, limited by how the VRAM is exposed by OpenCL (ie very fast if it's just all mapped into PCIe). And if it was basically offering VRAM as PCIe memory, then this potentially means the VRAM can be used for certain RAM niche cases, like hugepages: some applications need large quantities of memory, plus a guarantee that it won't be evicted from RAM, and whose physical addresses can be resolved from userspace (eg DPDK, high-performance compute). If such a driver could offer special hugepages which are backed by VRAM, then those application could benefit.

    And at that point, on systems where the PCIe address space is unified with the system address space (eg x86), then it's entirely plausible to use VRAM as if it were hot-insertable memory, because both RAM and VRAM would occupy known regions within the system memory address space, and the existing MMU would control which processes can access what parts of PCIe-mapped-VRAM.

    Is it worth re-engineering the Linux kernel memory subsystem to support RAM over PCIe? Uh, who knows. Though I've always like the thought of DDR on PCIe cards. All technologies are doomed to reinvent PCIe, I think, said someone from Level1Techs.

  • source
  • parent
  • hideshow 1 child comment
  • load more comments (1 reply)
  • [–] 42 points 11 months ago (1 child)

    GPU prices are coming to earth

    https://lemmy.today/post/42588975

    Nvidia reportedly no longer supplying VRAM to its GPU board partners in response to memory crunch — rumor claims vendors will only get the die, forced to source memory on their own

    If that's true, I doubt that they're going to be coming to earth for long.

  • source
  • hideshow 1 child comment
  • [–] 68 points 11 months ago (4 children)

    Guess what, GPU prices are about to go back up because they rely on RAM.

  • source
  • hideshow 4 child comments
  • [+] -13 points 11 months ago (3 children)

    GPU is a processor, right? It is the gfx card that will increase in price

  • source
  • parent
  • hideshow 3 child comments
  • [–] 24 points 11 months ago (1 child)

    Where are gpu prices coming down? 5070ti is what 800? 3090 24gb is 800-900 still. They not coming down anytime soon.

  • source
  • hideshow 1 child comment
  • [–] 24 points 11 months ago (1 child)

    Not for long, GPU partners will have to source their own VRAM for manufacturing Nvidia now.

  • source
  • hideshow 1 child comment
  • [–] 17 points 11 months ago (4 children)

    ... I was thinking about buying a 96GB DDR5 kit from the local computer store a few weeks ago, but wasn't sure it was actually worth €700. Checked again now and the exact same product costs €1500. I guess that settles it, 32GB will have to be enough for the next couple of years then.

  • source
  • hideshow 4 child comments
  • [+] 4 points 11 months ago (3 children)
  • [–] 10 points 11 months ago (1 child)

    Part of it the thrill of the hunt. I've caught some great deals over the years stalking marketplace.

    Got .iso storage after chia crashed
    Got a 3090 after Bitcoin asics took over
    Got a 5900x when the X3D parts came out

    But I've never seen decent RAM for sale, only single sticks or slow kits.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 1 point 11 months ago

    I fully agree it's worth waiting for the thrill of the deal.

    The Chia crash was great. A local dude I found on Facebook marketplace bought a whole array worth of drives and never got around to even opening the packages for them.

    I scored a great local deal on a PC last week, but unfortunately the only weak point of it is the 16GB of DDR5. So I guess for a while I'll just be chilling with equal amounts of GPU and system memory. Luckily I'm not a gamer and rarely limited by the system memory.

  • source
  • parent
  • [–] 1 point 11 months ago

    I just wanted to test if it was viable to run larger MoE LLMs on CPU, e.g. Qwen3-next-80B-A3B.. Even if I got acceptable generation speeds I'd probably get bored with it after a few hours, as with other local models. Had I got it for €700 it was pretty low value for money anyway, since my current RAM is enough for everything else I use the computer for. On the positive side, I can put that money towards a Steam Frame instead.

  • source
  • parent
  • [–] 11 points 11 months ago

    So.... the models are all trained up and now they need to run them is what I'm reading.

    You need lots of vram to train a model.

    An llm once trained can be run in much less vram and a lot of ram

  • source
  • [–] 30 points 11 months ago (18 children)

    This is very unfortunate, about a year ago I built my PC and only put in 32 GB of Ram, It was double I had on my laptop so I thought it should be enough for the beginning and I could buy more later.

    Already after 2 months I realizes I can do so much more because of the fast CPU in parallel but suddenly the amount of RAM became the bottleneck. When I looked at the RAM prices it didn't seem quite worth it and I waited. But that backfired because since then the prices never went down, only up.

  • source
  • hideshow 18 child comments
  • [–] 37 points 11 months ago (11 children)

    What are you running that needs more than 32Gb? I'm only just barely being bottlenecked by my 24Gb when running games at 4k

  • source
  • parent
  • hideshow 11 child comments
  • [–] 7 points 11 months ago (2 children)

    Two browsers full of tabs but that is not a problem, but once I start compiling AOSP (which I sometimes want to do for work at home instead in the cloud because it's easier and faster to debugg) then it eats up all the RAM imediatelly and I have to give it 40 more GB or swap and then this swapping is the bottleneck. Once that is running the computer can't really do anything else, even the browser struggles.

  • source
  • parent
  • hideshow 2 child comments
  • [–] 6 points 11 months ago (7 children)

    AI or servers probably. I have 40gb and that's what I would need more ram for.

    I'm still salty because I had the idea of going cpu & ram sticks for AI inference literally days before the big AI companies. And my stupid ass didn't buy them in time before the prices skyrocketed. Fuck me I guess.

  • source
  • parent
  • hideshow 7 child comments
  • [–] 6 points 11 months ago* (4 children)

    It does work, but it's not really fast. I upgraded to 96gb ddr4 from 32gb a year or so ago, and being able to play with the bigger models was fun, but it's not something I could do anything productive with it was so slow.

  • source
  • parent
  • hideshow 4 child comments
  • [–] 3 points 11 months ago* (1 child)

    You can have applications where wall clock tine time is not all that critical but large model size is valuable, or where a model is very sparse, so does little computation relative to the size of the model, but for the major applications, like today's generative AI chatbots, I think that that's correct.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 10 points 11 months ago

    I just had a look, 2nd of April I payed 67,000 KRW for one 16 GB stick, now the same one (XPG DDR5 PC5-48000 CL30 LANCER BLADE White), they only sell them in pairs, a pair costs 470,000 KRW in the same shop, so 235,000 KRW per 16 GB stick. That is a price increase of 250%, god damn.

  • source
  • parent
  • [–] 3 points 11 months ago*

    Last I looked, a few days ago on Google Shopping, you could still find some retailers that had stock of DDR5 (I was looking at 2x16GB, and you may want more than that) and hadn't jacked their prices up, but if you're going to buy, I would not wait longer, because if they haven't been cleaned out by now, I expect that they will be soon.

  • source
  • parent
  • [–] 2 points 11 months ago (3 children)

    Switch to Linux, double your available RAM for free.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 10 points 11 months ago (5 children)

    Buying used RAM on marketplace and hoping it isn’t broken. Hoping it was just stolen from a Best Buy. Fingers crossed y’all!

  • source
  • hideshow 5 child comments
  • [–] 14 points 11 months ago (4 children)

    I just recently bought a Samsung 16GB 5600MT/s stick for 45€ and received a 32GB stick instead! Sorry, but I wanted to brag x)

  • source
  • parent
  • hideshow 4 child comments
  • [–] 9 points 11 months ago (1 child)

    Prebuilts have become about the only way to get a deal for a while now.

  • source
  • hideshow 1 child comment
  • [–] 4 points 11 months ago

    Well that settles my idea of entertaining AM5.

  • source
  • [–] 8 points 11 months ago (2 children)
  • [–] 3 points 11 months ago*

    I just bought an Arc A310 for <$200. I couldn't find a set of 16x2 DDR4 in stock for less than that on Black Friday... Wild times for the RAM market.

  • source
  • 17 comments in c/hardware@lemmy.world view that post
    [–] 7 points 11 months ago* (4 children)

    You can finally get an RX 9070 (non XT), for under $600 MSRP now.

    NewEgg has the ASRock 9070 at $530 right now.

    You can even find 9070 XTs at $600.

    The next closest GPU from Nvidia in terms of performance per $ is the RTX 5060.

    Which is about half as performant as a normal 9070.

    The Nvidia 5070 TI is a bit more performant than a 9070, a bit less than a 9070 XT... and its at $750.

    https://bestvaluegpu.com/

    It would be extremely wild if a techtuber did a 'lets revisit price per performance' for GPUs right now.

  • source
  • hideshow 4 child comments
  • [–] [S] 1 point 11 months ago (3 children)

    Where I live the cheapest 9070 (non XT) costs USD 635, and this is sort of a special price (albeit it could become base in a few months). Median competitive price is closer to USD 700+.

    Cheapest 5060 (8GB) is around USD 340 (median competitive is close).

    5060 Ti (16 GB) is around USD 550 (median competitive is close).

    AMD in my country is no go outside of some older, super low end SKUs (but you can also get better Nvidia cards second hand).

    I don't care about brands, but where I live AMD is not really competitive (assuming like for like software support).

  • source
  • parent
  • hideshow 3 child comments
  • [–] 2 points 11 months ago (2 children)

    Ah, sorry, I should have specified I was talking about US prices. =X

  • source
  • parent
  • hideshow 2 child comments
  • [–] [S] 1 point 11 months ago* (1 child)

    No worries, I get that.

    I am just saying globally prices are more variable. I can't speak for EU specifically, but our prices generally track EU (more so than US).

    AMD needs to compete directly with Nvidia, but they won't since GPUs are a (US) government backed oligopoly.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 2 points 11 months ago*

    For the most part, I really do think AMD is trying to compete with Nvidia as best they can, in as many ways as they can, but Nvidia is simply utterly dominant and huge.

    They have very usable alternatives for almost everything Nvidia does, their cards are way, way easier to use on Linux due to being way more open with their drivers and such...

    Maybe, maybe, depending on exactly how the coming implosion of the US AI bubble plays out, they might be able to gain some ground against Nvidia... but that is an exceptionally complicated scenario to try to predict in any real detail.

  • source
  • parent
  • [–] 4 points 11 months ago (5 children)

    Of course I just bought a 970 xt for a lot a month ago, I did buy RAM before it shot up tho so thats nice. Slowly building a new PC thats not jank on Linux.

  • source
  • hideshow 5 child comments
  • [–] 5 points 11 months ago (4 children)

    Slowly building a new PC thats not jank on Linux.

    Sounds like you've done it. Unless you've got an unfotunate wifi card or something, still

  • source
  • parent
  • hideshow 4 child comments
  • [–] 3 points 11 months ago (3 children)

    Its literally just a GPU and the RAM. I need every other part as my kids are going to continue playing their smaller games on my current build like they do now. I can buy about 1 part a month thats under 200 dollars. I gotta sell one of these servers or something to get the CPU Idk lol I have a power edge r740 I dont really want to let go of but could.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 4 points 11 months ago* (1 child)

    This isn't a 'great' suggestion (and if you got a home server you're probably aware already), but there's a ton of cheap PCs right now that aren't Windows 11 compatible but are still relatively new.

    Might not be a bad deal to pick up an mid-tier business PC like an Optiplex or HP or whatever and 'upgrade' it with the GPU? As long as the power supply supports it, that is.

    I'm looking at one for a family member and its $125 for a pretty decent PC, more so once its running Linux.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 1 point 11 months ago

    I appreciate the suggestion lol I am going to take it slow though. I'll need time so I can also build a new desk to hold two desktops on too. Its a whole project in the game room that I dont make nearly enough money for. But life, uh, finds a way.

  • source
  • parent
  • [–] 1 point 11 months ago

    4080 still very pricey nearing $2k and the 5-series is even worse. 3080’s are in the $500+ range, so that’s decent.

  • source
  • [–] 1 point 11 months ago (4 children)

    GPU prices coming down are an indicator of the overall market sentiment around buying and building PCs with parts.

    5090 are still not anywhere at MSRP though reliably. Seems like 2800 is the lowest on a shelf near me. 9070xt comfortably under 600 now though.

    Memory prices are hitting the news everywhere but even at 200% or 300% of normal price you're still saving money over buying a gaming GPU in the first 6 months of the year when GPUs were chronically sold out. 16GB is kind of tight but 32gb is only $250 for 2x16gb @ 6000mhz CL30 at microcenter right near me. If this is "200-300%" pricing then I don't see what the big deal is.

    News outlets have pointed the finger at "AI" rollouts, but we're several years in to AI rollouts and although HBM memory used in AI cards has had elevated prices, it hasn't affected desktop memory very much until the last couple of months. I suspect pre-emptive hoarding by commercial system builders is more to blame as companies like lenovo have mentioned having 150%+ of normal part inventory levels. Buy and hoard has been a key strategy to handling taco tariffs too for all kinds of nonperishable products all year long.

  • source
  • hideshow 4 child comments
  • [–] 4 points 11 months ago* (last edited 11 months ago) (2 children)

    You've somehow gotten two downvotes, but nothing you've really said is incorrect.

    Well, I mean, I think $250 for 32gb sys RAM is a bit ridiculous, buuuuut...

    Yeah, yeah, for most GPUs, at least in the US?

    You scammed yourself if you bought one late last year / early this year.

    Nobody (in terms of retailers and GPU mfgers) expected the economy to fucking nuke itself after more or less Trump decided to turn a very, very shaky, almost/maybe "recovery" (really more like 'stabilization'), into a tumbling collapse down a cliffside.

    People also just arent't spending anywhere near as much on video games, the things you use the consumer grade GPU to play.

    (Expect a lot of AAA publishers/studios to either go very, very big into debt/buyouts, or just utterly collapse in 2026.)

    So, the result is, for the large part... demand has cratered, retailers are sitting on too much supply, and storing shit in a warehouse indefinitely costs money, so they've just been ticking down prices to try and induce demand.

    Consumer spending in general is collapsing, so uh... yeah, fairly good time to buy now, assuming you're not in the ~60% of Americans who do not have $1,000 set aside for an emergency.

    I did not know lenovo just straight up said they're sitting on a 150% of regular parts inventory, but uh yeah, if the entire computer industry basicslly becomes hoarders, yeah, yeah, that'd cause a price spike in system ram alright...

    It really is a simple as... system ram is modular and can go into a bunch of shit, GPU ram is already part of the board, and just generally, consumer oriented GPUs are waaaaay more niche than consumer oriented system RAM.

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

    I saw the lenovo thing on slashdot. I would link the articles directly but they all have paywalls for me https://it.slashdot.org/story/25/11/24/154202/lenovo-stockpiling-pc-memory-due-to-unprecedented-ai-squeeze

    The tariff thing has been astounding. The amount of announcements, exceptions and cancellations has made it nearly impossible to keep up. I think we just exempted brazilian beef and some other thing? Plus I believe semiconductors still have total exemption- sure maybe a GPU has a buck of tariffs on the aluminum in it but that's nothing lol.

    I had friends looking to build computers 6 months ago and balking at GPU prices. The prices overall today are just cheaper, unless you were not buying a GPU. CPU prices have come down a tiny bit too.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 4 points 11 months ago*

    I... am an econometrician, programmer and hardcore gamer, rofl.

    Yeah, watching this has been fucking insane.

    I used to work for a major international logistics company, made the reports for C Suite.

    I guarantee you people at my old firm were literally pulling out their hair when the de minimis exemption got annulled.

    ... anyway, as far as PC builds go?

    Here's the best possible bang for your buck, as far as I can tell:

    Minisforum BD790i se, + RX 9070.

    Put that, a 600W PSU, 32 gb sys ram, 1 or 2 TB SSD, and a big ole fan on top of the cpu, all inside an ITX lunchbox?

    ~$1500, and that'll get you over 60 fps on nearly anything, at nearly any settings level, at 1440p. Most games you'll be over 90 or 120fps.

    Its more or less a Steam Machine on steroids.

    ... presuming you are running Linux, lol.

  • source
  • parent