What a term! We explored so much about text to image AIs, however, we must not ignore the environmental impact of AI training. As we discussed in class and attempted during our assignment, LORA training requires a lot of energy even with advanced GPU. So I’d like to present a mini exploration on the power consumption and cost of training a LORA on the PonyDiffusion XL v6 model.
Data
GPU: AMD gfx_1030 16GB
CPU: Ryzen 9 5900X
Model: PonyDiffusion XL v6
Training Steps: 2000
Power Consumption:
Before Training: 351.688 kWh
After Training: 351.899 kWh
Total Used: 0.211 kWh
In New Jersey electricity price per kWh: $0.20/kWh (estimated with other fees)
So…
I trained a LORA on the PonyDiffusion XL v6 model for 2000 steps and measured the power consumption from my Home Assistant control panel. The model went from using 351.688 kWh to 351.899 kWh, an increase of 0.211 kWh. At $0.20 per kWh, that’s an extra $0.0422 in electricity costs. So for 2000 steps, the training used $0.0422 worth of power. That breaks down to $0.0422 / 2000 = $0.000021 per step, which is pretty cheap! The AMD RX6900XT and Ryzen 9 5900X handled it all efficiently. The low cost per step makes LORA training an affordable way to fine-tune diffusion models for specific outputs.
More Numbers… Just for Fun
How many LoRAs I could fine-tune with 1 dollar?
As stated in the last paragraph, each step costs $0.000021, which is 0.0021 cents each step. With a budget of 1 USD, I could run a whopping 47393 steps(which really is a lot of steps). Assuming each fine-tune runs 1600 to prevent overfitting, this comes down to just a little shy of 30 fine-tunes.
Leave a Reply