llama.cpp tensor split boosts dual-GPU speed

This is the kind of “free gains” benchmark that changes hardware advice fast: once tensor split is stable enough, the best upgrade path for local LLMs may be adding a second card instead of chasing a bigger single GPU.

• –The jump from 1580/44 t/s on one 3090 Ti to 2047/58 t/s on two cards is meaningful, especially for prompt-heavy workloads
• –`-sm tensor` appears to outperform older layer-splitting behavior on this setup, which matters for users trying to scale past one GPU
• –The result is from a mainstream `llama.cpp` build, so the optimization is moving from experimental patch territory toward practical default-adjacent usage
• –The gains are still workload-dependent, but the benchmark shows multi-GPU inference can now scale without needing datacenter-class hardware
• –For LocalLLaMA users, this is a strong signal that consumer dual-GPU rigs are getting better value from the software stack itself, not just from raw VRAM