Best GPU for running DeepSeek-V3 locally?

Question

honestly fed up with paying monthly for api credits just to have the service go down or get censored right when im in the zone for my coding project. im finally ready to just build a local box and deepseek-v3 looks like the dream but my current setup is basically a paperweight trying to run anything that big. i have a budget of roughly $2500 and im hoping to buy everything by friday so i can spend the weekend tinkering. what gpu is actually gonna handle this beast without me waiting five minutes for a reply? is a 4090 enough or do i need to hunt for dual cards to get enough vram to make it usable?

yzqjvpzvur · Accepted Answer

tbh a single NVIDIA GeForce RTX 4090 24GB is great for most things, but deepseek-v3 is a total vram hog. 24gb is gonna feel real cramped if you want to run anything other than the smallest quants. since youre on a $2500 budget for the whole rig, dont blow $1800 on one card. i usually suggest people look for two used NVIDIA GeForce RTX 3090 24GB cards instead. you can find them for around $700-$800 if you look around. having 48gb vram makes a world of difference for these huge models. just make sure you grab a solid psu like the EVGA SuperNOVA 1300 G+ 1300W because dual cards eat power for breakfast. spend the leftover cash on at least 128gb of system ram just in case you need to offload some layers to the cpu. its the most logical way to stretch your dollar for an llm build right now.

Dizaynersk_absl · Answer

@Reply #1 - good point! Honestly, it is pretty disappointing how much hardware this model eats. Even with your $2500 budget, trying to run the full DeepSeek-V3 is gonna be a real struggle... unfortunately a single 4090 just doesnt cut it for a 671B parameter beast. You really need total VRAM capacity more than raw clock speed for this specific project. To keep the whole rig under your budget, you are basically forced into the used market for GPUs:

Hunt for two used NVIDIA GeForce RTX 3090 24GB GDDR6X cards. You can usually find them for around $700-800 each on marketplaces.

48GB total VRAM lets you run decent quants, but youll still need Crucial 128GB DDR5 5600MHz RAM to offload the layers that wont fit.

Definitely grab a EVGA SuperNova 1600 G+ 1600W PSU because dual 3090s pull massive power under load. It is not as fast as the cloud setups, but it gets you away from the censorship. Happy to help you tweak the component list if you find some deals!