C.W.K.
Stream
Lesson 07 of 12 · published

Scaling Laws Today: Beyond Chinchilla

~10 min · scaling-laws, chinchilla

Level 0Token
0 XP0/94 lessons0/10 achievements
0/120 XP to next level120 XP to go0% complete

Scaling laws describe how loss decreases as you scale parameters, data, and compute. Two foundational papers:

  • Kaplan et al. (2020). Loss follows power laws in N (parameters), D (training tokens), and C (compute). Originally suggested allocating ~73% of compute to params, ~27% to data — leading to large, undertrained models like GPT-3.
  • Chinchilla (Hoffmann et al., 2022). Updated: scale params and data together. Compute-optimal ratio: ~20 training tokens per parameter. Many older models (GPT-3 included) were severely undertrained.

What changed after Chinchilla

Production-oriented teams realized training compute is paid once but inference compute is paid forever. So they over-train smaller models for cheaper serving:

ModelParamsTokensTokens / param
GPT-3 (2020)175B300B1.7 (severely under-trained)
Chinchilla optimum~20
Llama 2 (2023)7B2T~286
Llama 3 (2024)8B15T~1,875
Llama 3.1 (2024)70B15T~214

Llama 3 8B at 1,875 tokens/param is ~94× past Chinchilla-optimal. The "wasted" training compute is recouped many times over by the cheaper inference of a smaller model serving billions of tokens.

Code

Tokens-per-param sanity check·python
models = [
    ('GPT-3',     175e9, 300e9),
    ('Llama 1 7B', 7e9, 1.4e12),
    ('Llama 3 8B', 8e9, 15e12),
    ('Llama 3.1 70B', 70e9, 15e12),
    ('Llama 3.1 405B', 405e9, 15.6e12),
    ('DeepSeek-V3 (671B total)', 37e9, 14.8e12),  # active params
]
for name, p, t in models:
    ratio = t / p
    print(f"{name:>30}  tokens/param = {ratio:>7.1f}  (Chinchilla: 20)")

External links

Exercise

Pick a model you'd consider serving for 100M queries per month. Compute (a) one-time training compute (rough), (b) annual inference compute (per-query × queries × 12). When does the over-training premium pay back? Should you prefer a smaller, more-trained model or a larger Chinchilla-optimal one?

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.