Tesla M40 (Maxwell) has no INT8 DP4A - slower than the P40 it barely undercuts

M40 24GB at ~$189 looks like the cheapest 24GB card, but llama.cpp falls back to the FP32/CUDA-core path and it is SLOWER than a P40 despite similar VRAM.

Maxwell CC5.2 predates both INT8 DP4A and tensor cores; GGUF Q4/Q8 INT8 path cannot engage, so the card decodes on FP32 CUDA cores.

Only buy the M40 if ~$189 is truly the ceiling. Stretch to a P40 (~$110-280) for the DP4A path.