RTX 4000 Pro Blackwell SFF performance hacks (piszczek.pl)

Recent performance experiments on the NVIDIA RTX PRO 4000 Blackwell SFF have yielded significant improvements in AI model token generation rates, specifically for Qwen3.8 27B and Qwen3.6 35B A3B. By modifying the power limit from 70W to 120W and optimizing memory and core clock settings, the token generation rates increased by up to 67% for Qwen3.8 and 36% for Qwen3.6. These advancements highlight the potential for tuning configurations to maximize the efficiency of different AI models, underlining the complex interplay between hardware settings and AI workload demands. The experiment involved careful adjustments to power limits and memory frequencies, utilizing a modified VBIOS rather than traditional overclocking methods. The findings emphasize a nuanced understanding of GPU performance, showing that while increasing power can boost performance, the optimal settings differ by model and task. This research could lead to better resource management and configuration strategies for running AI applications on GPUs, making it a valuable addition to the toolkit of AI developers and researchers looking to harness the capabilities of cutting-edge hardware.