Optimus Alpha
Let's walk through an estimation of how many FLOPs were required to train GPT-3, using only reasoning and not searching for the exact numbers.
Step 1: Understand Model Size
GPT-3 (the original) had 175 billion parameters.
Step 2: Estimate FLOPs per Forward/Backward Pass
- For transformer models, a rough rule of thumb is that a forward-and-backward pass for a parameter takes ~2 FLOPs per parameter per token (1 for forward, 1 for backward).
- Each token processed in a batch involves the whole model.
So, FLOPs per token = 2 × number of parameters.
Step 3: Estimate Total Number of Tokens Used in Training
- GPT-3 was reportedly trained on about 300 billion tokens.
- Let's use 300B tokens as our assumption.
Step 4: Calculate Total Number of FLOPs
Total FLOPs = FLOPs per token × number of tokens processed
- FLOPs per token = 2 × 175B = 350 billion FLOPs per token
- Number of tokens = 300 billion
Now, multiply:
[ \text{Total FLOPs} = 350,\text{B} \times 300,\text{B} = (3.5 \times 10^{11}) \times (3 \times 10^{11}) = 1.05 \times 10^{23} ]
So,
[ \boxed{ \approx 1 \times 10^{23} \text{ FLOPs} } ]
Step 5: Check and discussion
This estimate is rough, but matches the ballpark for very large models. (Other nuances, like multiple epochs or data repetition, may factor in, but using "number of tokens processed" as above is standard.)
Final Answer
It takes about ( 1 \times 10^{23} ) FLOPs to train GPT-3.
This comes from:
- ( 175 ) billion parameters
- Each token: ( 2 \times 175 ) billion = ( 350 ) billion FLOPs
- ( 300 ) billion training tokens
- ( 350 ) billion × ( 300 ) billion = ( 1 \times 10^{23} ) FLOPs
Assumptions:
- Each forward+backward pass needs ≈ 2×params FLOPs/token.
- 300B tokens in total.
Notes:
- Reality varies with model and optimization, but this is a widely accepted estimating method.
