Comment by bean469

> comparable to mainstream NVIDIA GPUs

By this they probably mean RTX series GPUs? If so, then they are not comparing the hardware efficiency with the A100 / H100, etc. that are commonly used for training models