>NVIDIA cards are unreasonably expensive, you should buy AMD>dUdE bUt wAt 'bOuT mUh CUDA?????Oh yeah, and what are you going to do on your 8/12/16 GiB VRAM cards? Running local AI? Lmao.
Blender.
>>109552126what about my 32gb and 96gb cards?
>>109552141>completely missed the point
>>1095521268gb is enough for most image gen models, 12gb is enough for gemma 4 12b and 26b a4b if you offload inactive experts to system ram.AI on AMD is a dumpster fire because no dev optimizes for anything but cuda and zluda still isn't here to fix it, but AMD is definitely the choice if you want to game under linux because nvidia loses 30% or so gaming performance vs windows cause they only optimize their linux driver for AI.Personally I own both, neither brand is your friend and both hate you+want to take as much of your money as possible so just buy what works for what you want to do with it and performs the best for the excruciatingly high cost.
>>109552126Yes, what are you going to run with your ayyMD card VRAM besides delaying reboot due to memory leaks or unoptimized shit that either has memory leaks or gets relayed to the CPU because no support from AMD. >B-but AMD said they were working on itAMD promises, NVIDIA releases. But keep your toys, enjoy them, use it on your gaymes, live your childhood to your fullest.
>>109552145im not poor so your point is pointless
>>109552178that vram leak doesnt even work anymore
>>109552154all true except windows vs linux gayming perf.Was true during the x11 saga, today I see either no difference or slight bump vs windows on Fedora with nvidia 4000 series.
>>109552247Nvidia is still 30% slower in all VKD3D games. It's only DXVK where performance is similar to Windows but those are all old legacy games aka DX9.
>>109552178>Windowsnot my problem.>>109552154>8gb is enough for most image gen models, and 26b a4b if you offload inactive experts to system ram.Lol wtf are you talking about? 8 GiB is barely enough for Z-Image Turbo and if you wish you use something more advanced like Qwen Edit it's requires 24+ GiB (not to mention that generated images are still low quality).>12gb is enough for gemma 4 12bNot it's not, even for low quant models you need some headroom, and if you're using vision models it will require even more. And low quant are shit, you require at least Q4 to get decent results.>and 26b a4b if you offload inactive experts to system ram.This will still tank tops performance to unusable levelsReal useful shit starts from this https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF and it's already beyound reach for 16 GiB VRAM cards>AI on AMD is a dumpster fire because no dev optimizes for anything but cuda and zluda still isn't here to fix itNo it's fucking not. Torch rocm backend just works (now), llama-cpp and vllm also work, and with 16 GiB VRAM card you don't need more.
>>109552126I have a 5090 and 4080. 16gb is plenty of headroom of plenty of cool AI shit, and there are always workflows optimised for poorer people with older cards or cards with less VRAM.AMD cards specifically arent as good for AI stuff. I wish it wasn't so, but thats how it is
>>1095522778GB will run anima and illustrious, this is an anime website bloated 3DPD models need not apply.I get 60 tokens a second text gen with gemmy 12b q4_k_m on a single 3060 12gb and even more on 26b a4b, not sure what you're doing wrong if you're not a windows user.>DavidAUOh you're just retarded, I'm sorry for your mother's loss.
if you don't have at least a 16gb nvidia card you should be shamed off of this site for being poor tbdesuit's nobody's fault but yours for not having gotten one when the prices weren't as high as they are now
>>109552126>Running local AIas an adult with intact dick, no colored hair and who drinks only milk that comes from cows, I struggle to find any other use-case for external GPUs
>>109552154>8gb is enough for most image gen models>>109552587>8GB will run anima and illustriousDo you know if it's enough for Western animation images (and/or video) and LoRa training? Something like this. It's a Civitai LoRa (WAI-illustrious-SDXL v11.0).
>>109552606My 1080Ti tells you to fek off
>>109553221/ldg/ probably knows, I've never tried training a lora personally, but I've used loras with my 8gb tesla P4 fine before I upgraded to the 3060s.
>>109552261I literally get the same fps on windows 10 and fedora while playing Satisfactory, Elden Ring, TWW3 and Expedition 33 - which are not dx9 games. So i'm thinking troll or skill issue
>>109552982hardware transcoding is cool, which is interesting because i do not like trannies nor coding
>>109552606>doing only segmented OCR>have a regular 4070 with 12GB VRAMThis is all I fucking need, nigger. My 4070 does everything absolutely fine and since my own programm segments the pictures by itself and lets AI only OCR the segmented snipets it gets mostly everything correctly within the first try and doesn't even hallucinate because of muh context.
>>109552126what year is this post from? wtf is going on here?
>>109552178amd hardware works best on drivers not written by amd
Daily reminder that AMD released a 4GB card in YEAR OF OUR LORD 2026 and OP IS A FAGGOT
>>109552189Your spending habits indicate otherwise. :^)
>>109553401Why are you lying bitch? Every outlet proves you wrong.
>>109552585>>109552606>>109553221>>109553527Windows Retard Thread
I don't understand why autists here make everything a tribal brandwhore thing instead of whatever best suits your needs. Nvidia, AMD, Apple, Windows, Linux, they all have their pros and cons that depend on your use case.Fuckin autists man
>>109552126>AMD is cheaperAMD isn't even trying to be cheaper despite being worse at everything
12GB vram is plenty for videos with minimax h3. Nice shill attempt though
>>109552126>buy max VRAM>run AI to create a custom kernel to run the AI model of your choice
>>109552126I run chat/image/video models just fine on my 16GB card. Plenty of people run models on 8GB cards.
>>109554889Not him but I also get performance within margin of error difference between linux and windows with my 5070Ti. Either you're the one lying or your info is outdated
>>109552277>bad opinions>loves sloptunesIt checks out.
>>109557499probably amdrone who just hecking loves his favorite corpo, and other corpo bad
>>109552126sure feels good to have a 24gb 7900xtx.
