If you want 12GB of video memory for local AI on a budget, two cards to consider are Intel's Arc B580 and NVIDIA's older GeForce RTX 3060 12GB. Forge PC hasn't benchmarked either card, so there are no tokens-per-second or image-generation figures here. This comparison uses Intel's and NVIDIA's specification pages and the documentation for Ollama, llama.cpp and ComfyUI.
Specs side by side
Spec | Intel Arc B580 | NVIDIA GeForce RTX 3060 |
|---|
Architecture | Battlemage (Intel lists Q4 2024 launch), TSMC N5 | Ampere |
Memory | 12GB GDDR6 | 12GB GDDR6 (an 8GB version also exists) |
Memory bus | 192-bit | 192-bit (128-bit on the 8GB version) |
Memory bandwidth | 456GB/s (19Gbps) | Not listed on NVIDIA's page |
AI hardware | 160 XMX engines, 233 peak INT8 TOPS | 3rd-gen Tensor Cores, 3,584 CUDA cores |
Board power | 190W TBP | 170W graphics card power |
Recommended PSU | 600W | 550W |
Interface | PCIe 4.0 x8 | PCIe 4.0 |
Memory: same capacity, different bandwidth
Both have 12GB, so they'll fit models of a similar size. Our LLM quantisation guide explains how model size and quantisation decide what fits. Intel rates the B580's memory bandwidth at 456GB/s. NVIDIA's RTX 3060 page lists the memory size and 192-bit bus but not a bandwidth figure, so we can't put the two side by side from official pages alone.
Watch out for the 8GB RTX 3060. NVIDIA's page lists both 12GB and 8GB configurations, and the 8GB card has a narrower 128-bit bus. For local AI, make sure the listing says 12GB.
Software support
This is where the two differ most.
Ollama: Ollama's GPU documentation lists the RTX 3060 among supported NVIDIA cards (compute capability 8.6). Ollama provides additional GPU support through Vulkan, which it says is enabled by default when the back end is installed, and its Vulkan section links Intel's Linux GPU driver instructions; on Windows, most vendors' drivers include Vulkan support without extra steps.
llama.cpp: its SYCL back end documentation lists the Arc B580 as a supported Intel Arc B-series card. It also documents a workaround for B-series cards, where the driver can accept allocations larger than VRAM and move the overflow to system RAM, which can exhaust host memory when a model doesn't fit.
ComfyUI: NVIDIA cards use CUDA builds of PyTorch; for Intel Arc, ComfyUI's install guide uses PyTorch with xpu support. ComfyUI also offers a separate Intel portable build for Windows.
In short, the RTX 3060 uses CUDA, which all three tools support directly, while the Arc B580 is supported but takes a different back end in each app. Our guide to running local AI on an Intel Arc GPU walks through the setup.
Power and fitting
The B580 has a 190W board power rating and Intel recommends a 600W PSU with one 8-pin connector for its reference card; NVIDIA lists 170W and a 550W system for the RTX 3060. Partner cards vary, so check the specific model.
Which should you buy?
Choose the Intel Arc B580 if you want a newer card with Intel's rated 456GB/s of bandwidth and you're happy to set up Vulkan or SYCL back ends. It's also the newer gaming card; see Arc B580 vs RTX 5060.
Choose the RTX 3060 12GB if you want the CUDA route with the least setup in Ollama, llama.cpp and ComfyUI, or you're buying second-hand. Our used GPU for local AI guide covers what to check, and see our report on the RTX 50 shortage and the RTX 3060 in the UK.
For more options, see our best GPU for local LLMs.
Our verdict
Both give you 12GB for local AI. The RTX 3060 12GB is the lower-effort choice because Ollama, llama.cpp and ComfyUI all support CUDA directly; the Arc B580 is the newer card with 456GB/s of rated bandwidth, but you'll use Vulkan, SYCL or PyTorch xpu instead. Whichever you choose, check it really is a 12GB card.