GPU Unlock on the RTX 4080 32GB, Tested: +78% BF16 GEMM, but LLMs Barely Speed Up
I tested the Windows UEFI version of GPU unlock (AD103_win_unlocker), which lifts GeForce's half-rate FP32-accumulate Tensor limit, on a modded RTX 4080 32GB with an AM5 PC. BF16 GEMM went from 100.71 to 179.42 TFLOPS (+78%) and TF32/FP8 gained 40-63%, but 125B MoE LLM prefill improved only 9-12% and decode speed and llama.cpp quantized models barely changed.