LocalLLaMA debates GPU rigs for agentic coding

This is exactly the kind of hardware calculus that local inference enthusiasts face as the GPU landscape shifts — and the 3090 vs. 5060 Ti debate reflects real uncertainty about whether software will catch up to newer silicon.

• –2x RTX 3090s offer 48GB total VRAM and proven memory bandwidth, currently the community default for running 30B+ models locally
• –4x RTX 5060 Ti would yield more VRAM headroom but multi-GPU inference with consumer cards remains software-limited; llama.cpp and vLLM tensor parallelism support is uneven
• –Mac Studio M3 64GB has the best perf-per-watt and unified memory makes large model loading seamless, but limits GPU compute headroom and ecosystem flexibility
• –Intel Arc and AMD Ryzen AI 395 (NPU/iGPU) are fringe options with improving but still lagging software support
• –The "wait for software" bet on 5060 Ti is speculative — inference backends do evolve fast (GGUF multi-GPU, FlashAttention improvements) but timelines are hard to predict