Why isnt modding older GPUs to have more VRAM more of a thing now?Cant a 2090 or 3090 be upgraded to have like 60gb+ of VRAM?
>>110041673Most people don't have a soldering station, stencil, soldering equipment and memory chips aren't cheap.
What's the big deal with companies trying to make CPUs smaller and more compact when they could make them bigger and fit more transistors with ease?
>>110041673theres two things that stand in your waylimitations in the firmwareand the physical configuration of the pcbanf ofc skills, equiplent
It is, though it's primarily done by the Chinese. Alongside soldering on more VRAM chips, you also have to program the VBIOS.
>>110041686It has to be more than that though. I havent seen someone do testing/benchmarking on one of those chinese 96gb 5090s, but the cost of everything sky-rocket I would have thought would make this more popular
>>110041690yield.the bigger the chip, the more probable is it that you're gonna have a defect on itthink of it this way: you have a 100x100 grid of cellslets say its a given one out of 10 of them will be defectiveif your cpu is big, as in it takes lets say 10x10 cells, its expected that none of them will be defect-lessyoure gonna have a 0% yieldif your cpu's size is 1 cell, then you keep the good ones, and reject the defects, which should give you 90% yield
>>110041692>firmware limitationssolved with AI>physical configuration, skills, equipmentthe only real blockers
>>110041729>>firmware limitations>solved with AImentally handicapped take
>>110041729>solved with AIPost (a novel) modded firmware made by you via AI right now or else you're a faggot.
I want my 5070 ti modded to 32GB VRAM.
>>110041673It takes very good soldering skills, uncommon equipment (you're not going to be replacing RAM chips with a basic bitch soldering iron) and the extra VRAM itself. The latter here especially isn't going to be cheap or easy to find I'd guess, since it's not something that was ever sold to the general public outside of already assembled products.>>110041690>when they couldThe deal is they cannot. When they make a wafer, there are going to be some defects on it. If the CPU is large, then you can fit fewer CPUs per wafer and a defect in any of them might make the whole CPU, thus a large chunk of the wafer, useless. The larger the CPU, the harder it is to actually produce due to defects, as the area increases so does the chance of having a defect.Meanwhile if the CPUs (or CPU components) are small, a defect only invalidates a small chunk of the wafer and the chance per component of being impacted by a defect is smaller. You can just remove the small broker parts and use the rest. It's much easier to make small chips, from a practical manufacturing point of view.
>>110041673>>110041692>>110041729furthermore the real issue is the limitation of the chipsmost GPUs have 6 to 8 slots for chips but the highest you will have each chip at is like 16 Gb because nobody is fucking making any bigger than thatHBM chips are a bitch with firmware but worth a shot I guess
>>110041742>>110041737snailcats need not reply
>>110041742i do not have a modified gpu to test with nor a claude subscription>>110041737what makes you think this specifically is off the table?
>>110041673Well it didn't really make financial sense, until now perhaps. Actually get more VRAM onto the GPU is generally possible, however the issue is with the memory controller and having the card actually using the additional VRAM. There has been success in the past, but generally you only get to the stage where the cards recognizes the VRAM, but doesn't use it.It seems like the easiest DYI is getting shitty version, where they lazily market segmented, IE, not a design thing. So the same hardware and software accept because its literally designed to take more RAM.Of course the actual answer is because they want to exploit you. Way in the past it wasn't such a hurdle, but over time they down their cards, especially Nvidia (dont buy nvidia).Actually, I checked the prices for the VRAM a few years back and for slightly older cards the chips weren't expensive. Vram straight down the toilet cause they cuck us.The memory controller might have become less programmable and generally these things have become more integrated as with all PCB design. So, there might be some reasons for this. Creating a modular GPU is completely viable, though perhaps not profitable?
>>110041752hbm chips are not compatible with gddrdifferent packaging, even different controllers apparently
>>110041751Why don't they just arrange 4 good working chips in a grid within the CPU?
>>110041758>snailcatyoure mentally handicapped>>110041760i know its off the table because i saw demos of disassembling, then modifying firmware using claude with infinite tokensand, yeah, lol. fucking forget about itas to why? its the limitations of shatbotsthey re approximation machines so they accomplish tasks approximately, an approximate number of times
well you void the warranty for starters
>>110041673>moddingmost people in tech have no electronics or soldering experience whatsoever
The chinese have been doing this since the 2080ti.
>>110041751they shouldn't make defective chipstighten that shit up
>>110041760Just have some AI agent download a firmware dump from some firmware modding forum's anus, and have it apply a patch to it. You can use one of your old GPUs, dump the firmware, and apply a patch that limits the VRAM as a proof of concept.
>>110041808seems like a lot of work for me for no real benefit
>>110041808>Just have some AI agent download a firmware dump from some firmware modding forum's anus, and have it apply a patch to it.lolthe cargo cultist and his hallusinationspost the other balding faggot and his opengl port that lacks 3/4 of the features, and runs quake 3 at 60 fps on a modern machine
>>110041789>youre mentally handicappedpsychological projection is a tool of snailcats
>>110041855nonoyoure an grown ass man who believes in the equivalent of santa clausyoure just developmentally retardedarrested development kinda deal, not just- stupid
>>110041826I gave him an outline of a path to producing a modded firmware with aid (either partially, or entirely) from AI to show his argument's bullshit, you autist. Everyone's always making claims saying that X is solved and can be done by AI, yet that's all they are. Partially because it highlights the limits of AI, and the other half being because they're too poor to afford the tokens required for a project that AI would be able to accomplish.
>>110041673Because westerners, americans, aren't like the chinese or other thurdworlders who actually explore their hardware, no tinkering willing, everyone at mercy of what companies and manufacturers offers.
>>110041673While VRAM modding (soldering higher-capacity memory chips onto a graphics card) has become a popular niche hobby for cards like the RTX 3070 or RTX 2080, upgrading a flagship card like the RTX 3090 to 60GB+ is physically and architecturally impossible
>>110041696Chink days centers are doing exactly that and renting time on them.
>>110041862read up on integrated informational theory snailcat
>>110041696There's no proof that the 5090 96gb even existsThe two listings that went up had an abnormally low asking price and are probably scams
>>110042209>muh ai is le consciousyoure completely fucking retarded
