RTX 5060 Ti benchmark hub expands with dual-card recipes and static results explorer

A community benchmark project tracking RTX 5060 Ti local LLM performance has matured into a structured recipe and results hub. The club-5060ti repository now ships schema-validated benchmark JSON, a static results explorer, and documented inference recipes for both single-card and dual-card RTX 5060 Ti 16GB configurations. The project covers llama.cpp and vLLM workflows, with particular attention to Blackwell-specific features like NVFP4 and MTP that may not transfer cleanly to other GPU architectures.

Each benchmark entry logs exact hardware, runtime, model, quantization, context length, KV cache settings, token counts, prompt eval speed, generation speed, and caveats. The maintainer emphasizes that the value lies in the reporting discipline rather than universal performance claims — the numbers reflect a specific setup, and mixed-GPU or non-5060 Ti results should be reported as separate hardware lanes. The project has absorbed older llm-bench data as archived historical rows, with plans to rerun key cases under the new protocol rather than carry forward mixed-method results.

One early finding: llama.cpp with GGUF is the safer first test for non-5060 Ti or mixed-architecture setups, while vLLM's Blackwell-specific optimizations should not be assumed portable without validation. The maintainer is soliciting dual 5060 Ti results from different CPU and motherboard combinations, mixed-GPU and non-5060 Ti llama.cpp results, vLLM version drift reports, and clear failure logs alongside successful runs. The next useful contributions would be failure reports that document what doesn't work — version conflicts, out-of-memory crashes, and mixed-GPU edge cases — rather than only successful runs. That discipline is what separates a useful benchmark from a collection of anecdotes.