Compute-Optimal Allocation Between Model Size and Data Under a Parametric Scaling Law: Closed-Form Solution, Convexity Proof, Misallocation Penalty, and Sensitivity to the Fitted Constants
- 1Former Officer, Chaitanya Godavari Grameena Bank (CGGB)
Abstract
Under the parametric loss L(N,D) = E + A/N^α + B/D^β with training compute C = 6ND, I derive the compute-optimal model size and token count in closed form: N* = G(C/6)^{β/(α+β)} with G = (αA/(βB))^{1/(α+β)}. I prove that the constrained problem is strictly convex in log N, so the solution is unique, and derive a quadratic approximation to the loss from misallocation. Two published fits are evaluated: the rounded constants of Hoffmann et al. (2022) and the re-estimate of Besiroglu et al. (2024), both read directly from the papers' text. They imply different allocation rules. At the Gopher budget of 5.76 × 10²³ FLOPs, the first gives N* = 32 B parameters and 93 tokens per parameter; the second gives N* = 72 B and 18 tokens per parameter. The ratio drifts with compute in opposite directions under the two fits (50 → 153 against 22 → 16 tokens per parameter over 10²¹–10²⁶ FLOPs). Misallocation is cheap near the optimum: a model 2× too large raises loss by about 0.005 nats, and offsetting a 4× oversize needs 1.8–1.9× more compute. The prescriptive content of "compute-optimal" depends on second-decimal differences in fitted exponents.
Related
- Hidden Convexity in Multi-Retailer Cold-Chain Allocation: A Cholesky-Spectral Second-Order Cone Programming Approach
- Delay-Certified Readiness Barriers for Overactuated Multirotor Allocation Without Lipschitz Constants
- Parametric sensitivity assessment of slip flow and thermophoretic effects on a vertical stretching surface using response surface methodology
- An L-Shaped Sparse Array with Reduced Sensitivity to Parametric Mutual Coupling for 2D DOA Estimation
- Parametric sensitivity characterization of two-way time transfer over hollow-core fiber links up to 91 km
BibTeX
@misc{aguru2026computeoptimal,
title = {{Compute-Optimal Allocation Between Model Size and Data Under a Parametric Scaling Law: Closed-Form Solution, Convexity Proof, Misallocation Penalty, and Sensitivity to the Fitted Constants}},
author = {Manoj Kumar Aguru},
howpublished = {Zenodo},
year = {2026},
month = oct,
doi = {10.5281/zenodo.23450087},
url = {https://zenodo.org/records/23450087},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Compute-Optimal Allocation Between Model Size and Data Under a Parametric Scaling Law: Closed-Form Solution, Convexity Proof, Misallocation Penalty, and Sensitivity to the Fitted Constants
AU - Manoj Kumar Aguru
PY - 2026
DA - 2026/10/18
PB - Zenodo
DO - 10.5281/zenodo.23450087
UR - https://zenodo.org/records/23450087
AB - Under the parametric loss L(N,D) = E + A/N^α + B/D^β with training compute C = 6ND, I derive the compute-optimal model size and token count in closed form: N* = G(C/6)^{β/(α+β)} with G = (αA/(βB))^{1/(α+β)}. I prove that the constrained problem is strictly convex in log N, so the solution is unique, and derive a quadratic approximation to the loss from misallocation. Two published fits are evaluated: the rounded constants of Hoffmann et al. (2022) and the re-estimate of Besiroglu et al. (2024), both read directly from the papers' text. They imply different allocation rules. At the Gopher budget of 5.76 × 10²³ FLOPs, the first gives N* = 32 B parameters and 93 tokens per parameter; the second gives N* = 72 B and 18 tokens per parameter. The ratio drifts with compute in opposite directions under the two fits (50 → 153 against 22 → 16 tokens per parameter over 10²¹–10²⁶ FLOPs). Misallocation is cheap near the optimum: a model 2× too large raises loss by about 0.005 nats, and offsetting a 4× oversize needs 1.8–1.9× more compute. The prescriptive content of "compute-optimal" depends on second-decimal differences in fitted exponents.
ER -