Benchmark: GPU vs. Fly

Solving the AI Energy Problem

Here’s my Sunday morning thought experiment that I had to get out of my system: How does the modern advancement in AI stack up against nature? Recently, some doomsayers have been warning us that the singularity is around the corner. Can I put some numbers on this?

So, how do we benchmark machines against biological brains? Since I’m not going to set up a neuroscience and AI lab before lunch, the next best thing is to search the internet for existing experiments that are well-documented and follow the same general principles.

I chose reinforcement learning applied to a simple task: based on sensory input, take the correct path and receive a reward. With the aid of AI, I found two research papers that fit the bill.

1. Training Fruit Flies to find sugar while navigating a tubes structure – Rest Is Required to Learn an Appetitively-Reinforced Operant Task in Drosophila

2. Training AI agents how to fly in a simulated 3D space – LARGE BATCH SIMULATION FOR DEEP REINFORCEMENT LEARNING

Obviously, neither study was originally motivated by what I’m looking for; nevertheless, they contain the data I need. From both, I can extract a key metric: the time required for training. These fundamental data points are enough for a quick back-of-the-envelope calculation to estimate the energy required for these comparable tasks.

Results

It turns out that with the proper mix of training sessions and rest, it takes 2 to 3 days to train a fruit fly to take the correct turn and go down the tube with a sugary reward at the end. In contrast, a massive 64-GPU cluster requires 72 hours to train AI agents to recognize safe paths in 3D space.

So, practically the same time. Interesting. Energy analysis is pretty straight forward:

GPU:

Let’s use the spec for NVIDIA Tesla V100:

Maximum power consumption: 300W or 0.3 KW.

Daily consumption: 0.3 KW × 24 hours = 7.2 KWh.

64 of these bad boys: 64 × 7.2 KWh = 460.8 KWh

Running for 3 days: 3 × 460.8 KWh = 1,382.4 KWh

Fly:

Daily food consumption: 0.8 milligram sugar, or 0.0008 gram.

Sugar has 4 kcal per gram. Therefore 0.8 milligram => 0.0008 g *4 = 0.0032 kcal

Over the course of 3 days: 0.0032 kcal × 3 = 0.0096 kcal

Conversion to kWh:

1 kcal = 0.001162 kWh

0.0096 kcal × 0.001162 kWh/kcal = 0.0000111552 kWh

GPU’s energy / Fly’s energy = 1,382.4 / 0.0000111552 = 123,924,268.503

Simply put, GPU technology requires approximately 124 million times the energy needed to train a fly on similar tasks.

To visualize this, it's like comparing the distance from Earth to the Sun to the height of One World Trade Center.

This tells us that in order to scale meaningfully, our approach to computer hardware must radically change. We need to keep emulating nature. Harnessing the potential of biological brains remains in the realm of science fiction. The mentats from Dune may be the most well-known example. In Ex Machina, Ava has “wetware” in her skull. My favorite is Nightwings by Robert Silverberg—this novella is practically poetry.