Best value gpu right here.
>>109741953>3090 having 50% more vram for the same priceoh no no no no
>>109741953Probably 9070xt
>>109741968>getting a slower and hotter gpu because muh vramplease don't tell me you actually do this
>>109741992>slowerthe only reason to get a better gpu is for local ai, surely you're not playing modern goyslop games, right?so having 50% more vram and thus loading more of the model into the gpu is gonna make it much faster, and the 3090 already has slightly faster memory bandwidth.
>>109741992the connector on the 5070ti gets way hotter
>>109742003>ai>local aimother of all memes
>>109742040yes little goy, i actually do coding work with local ai and dont use my gpu to listen to UE5 leftshit black woman cinematic lectures.
>>109741953That would be the 9070 XT. The 5070ti is expensive and you won't need the proprietary features.
>>109742048You can do AI with the 9070XT. It's not an excuse anymore. AMD fixed their shit very quickly somehow.
>>109742067lol. lmao even. search the archives of lmg/ldg for AMD and see how good it is.
>>109742067AMD got supported a lot better for ai workloads, but even if and when it gets better it will not get on par with nvidia simply because now everyone that does ai does it on nvidia and so new developments happen there first..Don't get me wrong: I want nothing more than buying a gpu with more vram for less dollars..
>>109742086That's cool but if you can't deal with a bit of problems you shouldn't be doing AI in the first place.>>109742075Works for engineers.
>>109742075i use a 7900XTX; pytorch is the backing for 95% of local inference stuff and they got properly sorted for AMD backends. glm 5.3 flash can set up anything using pytorch like comfyui or forge for your specific GPU and ROCm version if you're too dum to into venv
3090 sisters we're losing not like this...
>>109741953buy an ad
>>109741953Just buy a fucking RTX 5080 and stop fucking around. RTX 5090 was extremely expensive last time I checked.
>>109742067>16gb vram vs 24gb 3090>640gb/s vs 940gb/s 3090 memory bandwidth
>>109742125Some 5080s were pretty "cheap" recently, I got a nice PNY one on sale for roughly Founders MSRP, cheaper than 5070Tis...
>>109742097What are your Qwen 3.8 27b numbers amdchad?
you need a blackwell 6000
>>109742003>Local AIThe 24gb vs 16gb will go a longer way desu
>>109741953>Best value gpu right here.ftfy
>>1097421733.7 tokens/sec. AMD won.
>>10974217324GB isn't really enough for qwen 27b in the first place, you should really sub to opencode go or openai plus or pro. my local GPU work is blender rendering and training smol neural networks w/ pytorch
