BEAST LAB / LOCAL AI GENERATION
How does changing the GPU configuration affect generation? Two test series use the same models, inputs and settings.
FRAME / 0001
01 / BENCHMARK
Cold-start load time is tracked separately. The speed comparison uses three warm runs per configuration.
02 / OUTPUT GALLERY
Images, video and audio from both configurations will appear here. This first image comes from a trial run.
TRIAL RUN / NOV 7, 2026
Krea 2 Turbo · INT8 · 1024 × 1024 · 8 steps
Another model was downloading to the same USB SSD, so this run is excluded from the comparison. ¹ NVML reading for the entire GPU.
03 / TEST RIG
Hardware and software verified after the November 7, 2026 reboot.
ASUS ProArt · 16 GB VRAM
320 W baseline power limit · Driver 615.71.098 cores / 16 threads
32 GB DDR5-6000 · ASUS ROG STRIX B650E-IMaestro 4.0.0 · PyTorch 2.10.0 + CUDA 13.0
Kernel 7.2.15-300.fc45.x86_64SanDisk Extreme Pro · Btrfs
Stores the Maestro model weights04 / PROTOCOL
Capture the model weights and version, prompt, seed if supported, input files, dimensions, steps, frame rate and duration.
Record the cold start separately, then measure three runs with background downloads paused and live previews disabled.
Compare mean time, variability, VRAM, power and temperature. For video and music, also compare generation time with output duration.
The GPU Unlock series depends on a compatibility review and approval to proceed. It uses the same settings as the baseline.