Dual 7900 XTX Build Targets Local LLMs

The hardware idea is plausible on paper, but the software and platform friction are the real story here: local inference usually cares more about VRAM capacity and runtime compatibility than perfect PCIe symmetry.

• –The bigger constraint is not just lane count; it is whether ROCm, llama.cpp, or ExLlamaV2 can reliably use both AMD cards across reboots without enumeration headaches
• –Asymmetric links may be tolerable for inference workloads, but tensor parallelism will still pay communication overhead, especially once model sizes or context windows grow
• –The M.2-to-GPU route is clever for lane-hungry ITX builds, but it adds another compatibility layer that can become the failure point before bandwidth does
• –If the goal is simply to run larger models locally, 48GB of aggregate VRAM is compelling; if the goal is predictable throughput, a single stronger GPU may be the safer bet
• –This is less a performance question than a systems-integration question: thermals, driver behavior, and runtime support will decide whether the build is elegant or fragile