[bevekspldnw]

VRAM in excess of the model one is using isn’t useful per se. My use cases require high throughput, and on many tasks the A6000 executes inference at 2x speed.